Bio inspired optimization techniques for disease detection in deep learning systems

Abstract

Numerous contemporary computer-aided disease detection methodologies predominantly depend on feature engineering techniques; yet, they possess several drawbacks, including the presence of redundant features and excessive time consumption. Conventional feature engineering necessitates considerable manual effort, resulting in issues from superfluous features that diminish the model’s performance potential. In contrast to recent effective deep-learning models, these may address these issues while concurrently obtaining and capturing intricate structures inside extensive medical image datasets. Deep learning models autonomously develop feature extraction abilities but require substantial computational resources and extensive datasets to yield significant abstraction methods. The dimensionality problem is a key challenge in healthcare research. Despite the hopeful advancements in illness identification with deep learning architectures in recent years, attaining high performance remains notably tough, particularly in scenarios with limited data or intricate feature spaces. This research endeavors to elucidate the integration of bio-inspired optimization techniques that improve disease diagnostics through deep learning models. The targeted feature selection of bio-inspired methods enhances computational efficiency and operational efficacy by minimizing model redundancy and computational costs, particularly when data availability is constrained. These algorithms employ natural selection and social behavior models to efficiently explore feature spaces, enhancing the robustness and generalizability of deep learning systems. This paper seeks to elucidate the efficacy of deep learning models in medical diagnostics by employing concepts and strategies derived from biological system ontologies, such as genetic algorithms, particle swarm optimization, ant colony optimization, artificial immune systems, and swarm intelligence. Bio-inspired methodologies have exhibited significant potential in addressing critical challenges in illness detection across many data types. It seeks to tackle the problem by creating bio-inspired optimization methods to enhance efficient and equitable deep learning for illness diagnosis. This work assists researchers in selecting the most effective bio-inspired algorithm for disease categorization, prediction, and the analysis of high-dimensional biomedical data.

Introduction

Biologically-inspired deep-learning models integrate design principles and approaches of deep-learning models based on their biological inspiration; bio-inspired deep-learning models offer potentially helpful methods of enhancing conventional deep-learning models. This is based on the capability and flexibility in nature that has inspired a model that tries to emulate communication, neural networks, and learning and behavior of different species of animals and human beings¹. It is a branch of artificial intelligence that adopts concepts from neuroscience, the idea of evolution², and swarm intelligence³, among others, to address the challenges of deep learning, such as retrofitting models to massive datasets and robustness. The intelligence procedures in various systems and machine technologies are also increasing. From the perspective of clinicians, optimization may focus more on the interpretability and clinical validity of the outcomes to guarantee that the insight generated is consistent with the domain expert’s knowledge and can be applied in clinical situations. Providers often want to consider individual cases of a particular disease and focus on the specific population and data set⁴. Further considerations for patients focus on the accessibility and user-friendliness of the approaches to spread the idea of using advanced diagnostics among people and provide them with more opportunities to control their health.

Humans are also an inspirational resource for the different aspects of bio-inspired deep learning⁵. One of the concepts developed to model the conditions and the operations of neural networks naturally upon finding their root in the structure and the functioning of neural networks. Bio-inspired techniques are, however, primarily focused on the innate capability to emulate the brain’s interconnected structure and functions. The networks have as their primary goal the replication of the spiking timing of the neurons and furthering computation efficiency in terms of time. Thus, they demonstrate how such models will gradually look and for what tasks, in particular, they may be helpful in robotics and brain-computer interfaces that imply real-time implementation. The efforts of designers and models circulating in the biological field may assist in advancing medical diagnostics by introducing bio-inspired techniques to develop more accurate and precise ideas and methods of disease diagnosis and treatment strategies. Dimension reduction, variable selection, and biomarker discovery are all significant applications of high-dimensional data in bioinformatics.

Besides works involving brain-inspired approaches in bio-inspired deep learning, various evolutionary and swarm intelligence⁶ algorithms are also being investigated. Genetic algorithms and evolutionary-based solutions like particle Swarm optimization are similar to the natural selection process in classifying the best architectures and parameters of deep learning⁷. It is a process of artificial evolution where a population of potential solutions are processed in iterations, and iterations of surviving solutions are combined to form new solutions. The possible application of bio-inspired deep learning models to solving challenging optimization challenges opens up opportunities for creating more intelligent and versatile AI technologies inspired by nature. It is impossible to provide all references to the 108 review papers and monographs on bio-inspired optimization algorithms we reviewed. Most contributed research articles contain short review sections – often only several pages long. The points explained above make us believe that we have developed a unique survey that covers several topics, which none of those mentioned above sources covered and that we are aware of can be found in one source. In other words, we tried to provide a combination of several objects that included the latest information and presented the broadest possible coverage of the time we could make. In the current paper, we attempted to prepare a thorough and comprehensive source of information on the main areas among numerous and often duplicative and, in some cases, even conflicting bio-inspired optimization algorithms⁸ that can be easily understood by the broad scientific audience coming from a variety of specializations. Using bio-inspired algorithms, including genetic algorithm (GA) and Particle swarm Optimization (PSO), improves deep learning model robustness and generalization performance, especially for high-dimensional biomedical data. Bio-inspired algorithms can identify essential features to decrease dimensionality while boosting model accuracy. Bio-inspired algorithms efficiently search within the hyperparameter space to discover optimal configurations that produce robust models.

Research approach

This paper presents an exhaustive literature survey from 2020 to 2024 related to nature motivated techniques (NMT) in neural disorders, heart disease, cancer, kidney disease, diabetes diagnosis, etc. This paper focuses on nature-inspired optimization strategies, disease identification, and bio-inspired methods with their application in disorder identification and prediction.

In this paper, three basic research queries (RQs) are answered such as.

• 1) What is the importance of bioinspired techniques in all such diagnoses?

• 2) Several roles played by bioinspired techniques in disease diagnosis?

• 3) Which latest bioinspired algorithms are applied to each disorder detection?

Various Nature Motivated Techniques (NMT) are utilized for disease discovery and classification to address these inquiries. Academic databases such as WOS, Scopus, PubMed, and PsycINFO serve as sources for information collection. Furthermore, the facilitated research articles concerning bioinspired approaches contribute to the acquisition of practical content. Keywords utilized for information retrieval include “Bio-inspired algorithms in disease diagnosis,” “biomimicry in brain disease diagnosis,” “heart disease diagnosis,” “Review/State-of-the-art,” among others. Multiple categories of research related to bio-inspired algorithms, including as ACO, ABC, and GA, have been recognized. Nature-inspired algorithms were introduced to the field of image processing in 1990. The current researcher has been observed to prefer NMT over various techniques in numerous applications.

Hierarchical classification of bioinspired optimization algorithm

Biologically informed deep neural networks are new deep learning models that take inspiration from biological systems like human brains or natural selection to improve conventional deep learning architectures⁹. Traditional feature engineering demands extensive manual work that creates problems from extra features, reducing model performance potential. Deep learning models automatically generate feature extraction capabilities yet need ample computational power alongside vast datasets to produce meaningful abstraction techniques. The consideration has been done without reference to particular applications of the algorithms in mind. These models can be interpreted as attempts to reproduce living organisms’ highly efficient information processing, learning, and adaptation mechanisms. The bio-inspired optimization algorithms, Genetic Algorithms, and Particle Swarm Optimization efficiently replace high-dimensional data by identifying the most significant features. Specific feature selection improves computational speed and operational performance by reducing model redundancy and computing expenses, mainly when data availability is limited. These algorithms use natural selection and social behavior models to search feature spaces efficiently because they make deep learning systems more robust and generalizable. The field of bio-inspired deep learning encompasses various approaches, each inspired by different aspects of biology:

Neural networks inspired by the brain

ANNs already loosely imitate the structure and functions of the brain. However, bio-inspired deep learning seeks out biological principles for developing biologically plausible neural network architectures, such as spiking neural networks (SNNs), which are very similar to the functioning of biological neurons and synapses¹⁰.

Evolutionary algorithms

Genetic algorithms are a class of evolutionary computing that uses natural selection to optimize deep learning models¹¹. They create and continually modify a group of potential solutions, using genes/parameters (e.g., neural network architectures, hyperparameters) from the previous generation to create new generations in an effort to create fitter individuals/models.

Swarm intelligence

Swarm intelligence sampling methods, such as particle swarm optimization and ant colony optimization, are based on the behavior of social insects or groups of birds. These methodologies are based on population-oriented evolutionary processes by which deep learning architectures are revamped using local interactions and knowledge-sharing mechanisms among the various agents in the population.

Membrane computing

Membrane computing theory is found on the biological membranes of the cell. More advanced systems tasks have been directed toward solving problems like pattern recognition, classification, and optimization in deep learning networks for more efficient computations and a more powerful information processing model¹².

Neuromorphic computing

Efficient information processing by the brain is suggested as Neuromorphic computing, where artificial hardware using various neural network architectures is developed to emulate brain-like capabilities. They include specialized hardware comprising neuromorphic chips that are low power and parallel to the processors to solve the deep training tasks with the least power consumption. There was a promise from biologically inspired deep learning methods to address some other issues in the traditional deep learning methods, including scalability, robustness, interpretability, and energy challenges¹³. The models remain adaptable since they enable better handling of incomplete and noisy biomedical data. The deep learning models can be characterized as a holistic biological approach for deep learning models. Thus, these disease detection methods in deep learning systems are considered a new and ethically significant way of machine learning built on bio-inspired machine learning approaches that improve disease detection capabilities. These techniques employ adaptation driven by biological concepts of evolution and ‘swarm intelligence’ as well as other AI techniques to optimize the architecture and parameters of the algorithms for improving the performance of disease diagnosis using the architectural deep learning approach. Here’s a detailed view of how these techniques are applied:

Genetic algorithms (GA)

The original version of the Genetic Algorithm (GA) was created in 1975 by John Holland of the University of Michigan. The genetic algorithm is an iterative search method that uses the theory of evolution and probability to solve the global optimization problem. This evolutionary technique is founded on Darwin’s theory of natural evolution, whereby the saying ‘survival of the fittest’ comes into play. The one says that the desirable organisms will be inherited while the undesirable ones will be destroyed. The variability of different people through generations can be termed natural evolution¹⁴. Therefore, the modifications that fit in the genes are maintained in successive generations. In natural transformation, some undertakings occur, such as a crossover of chromosomes, mutation of genes, and selection of the best genes for future generations. During GA execution, different genes are articulated as chromosomes expressed in string form. The string effectively has several properties that result in less or no interaction. It utilizes three leading operators: crossover, mutation, and selection processes. In the process of evolution, the old generation individuals are replaced by new individuals whose average fitness is higher than the old generation. Figure 1 shows the flow diagram of the Genetic Algorithm.

Fig. 1

Working mechanism of genetic algorithm.

Genetic algorithms are optimization algorithms based on the theory of natural selection and genetics. The application of GA in the disease detection model is in developing DL models based on the process of growing populations of candidates. A GA is considered a relative optimization algorithm. The genotype of chromosomes refers to these individuals and is a potential solution to the optimization problem. GA will repeat the process by populating potential model architectures and then using fitness functions to assess their performance. It uses a selection operator based on accuracy or sensitivity to choose the fittest individuals. It records the result of a mutation or crossover operator on the selected individuals to create the population for the next generation. Through exploring various model architectures and specific parameters, GA can aid in finding more suitable configurations for disease detection tasks.

Chromosomes are then selected from the population of the chromosomes based on their fitness. These are the chromosomes/individuals that are the fittest, and a subset is chosen from the population. The assessment of which individual should be the fittest can be done based on ranking, or some procedure may be selected to determine which individual is the fittest¹⁵. Such methods include roulette wheel selection and tournament selection. In this way, members of the population suggested are measured against each other, and the fittest of the lot are selected as potential parents for the next generation. In this way, the quality of solutions will increase for each generation. The crucial Darwinian steps – recombination and mutation – are carried out before each of these evaluations on the group of members chosen as the fittest.

Particle swarm optimization (PSO)

Another essential advantage of the PSO is that each particle maintains the location of the last good position known as P best (Personal best). PSO also has another distinguishing trait of information sharing from members of the swarm. Particle swarm optimization is based on population and inspired by swarming found in birds, fishes, etc¹⁶. Deep learning models can be trained using PSO for diseases in the detection class. Frontier maintains a population of candidate solutions (particles) that adjust their positions in the search space according to their own experience and the best performance of the swarm. Interactions and information diffusion among particles in the PSO allow for determining the optimal or near-optimal positions in the search space for disease detection models. The working mechanism of PSO is illustrated in Fig. 2.

Fig. 2

Working mechanism of particle swarm optimization.

Ant colony optimization (ACO)

A metaheuristic algorithm for optimization inspired by the foraging behavior of ants is developed¹⁷. It is understood that ants as a swarm determine the shortest paths from their colony to the food supplies and vice versa, using worker ants depositing their pheromone trails in their way and the rest of the workers following these trails. A particular ant (k) mimicking the behavior of ACO will choose a random path, and a route in the process will give rise to the graph and experience of pheromones at the graph edges¹⁸. The working mechanism of ACO is illustrated by reference to a very simple flowchart, as depicted in Fig. 3.

Fig. 3

Working mechanism of ant colony optimization algorithm.

ACO is a simulation of the foraging behavior of ants where many ants deposit pheromones to exchange information, which leads to the construction of shortest paths between food sources. ACO can be applied in disease detection to optimize the feature selection and extraction part of deep learning models¹⁹. ACO generates solutions in an iterative process in which the movements of the artificial ants are applied to the graph, representing the feature space. Ants place pheromone marks on the features depending on the role of this feature for the classification of disease instances, with higher pheromone levels denoting more compelling features. ACO can leverage the dynamic of the ant colony to identify features that add discriminative power to deep learning models used for disease diagnosis.

Grey Wolf optimization

The motivational idea of this technique is a way of hunting and hierarchy of leadership for gray wolves. The batch stays together, and the alpha is a group member expected to decide where to sleep and hunt. The batch’s second wolf is called beta, which helps in decision-making with alpha²⁰. Third, the smallest batch is called omega and has the task of collecting information from the wolves. (i) Chase after the prey, (ii) Buzz around and disturb the prey; (iii) Beat up the prey. In the hierarchy of the GWO, the one with the solution might be alpha. The last candidates will be considered as solutions to be omega²¹. GWO is becoming the preferred approach for researchers in searching for solutions because of its simplified procedure and fewer variables that need to be tuned to obtain global optima. The working mechanism of ACO is illustrated by reference to a very simple flowchart, as depicted in Fig. 4.

Fig. 4

Working mechanism of GWO algorithm.

Memetic algorithms

Memetic algorithms are a type of evolutionary computation that uses local search inspired by cultural evolution, as shown in Fig. 5. Regarding disease detection, memetic algorithms can fine-tune the parameters of deep learning models and optimize the development of feature representations^22,23. Memetic algorithms consist of a population of candidate solutions that employ genetic operations such as crossover and mutation. It includes genetic operators such as crossover, mutation, and local search strategies to complete the search space. The integration, exploration, and local exploitation by memetic algorithms will enhance the performance and robustness of the deep learning model in disease diagnosis.

Fig. 5

Working mechanism of memetic algorithm.

Typically, memetic algorithms consist of five steps:

• 1. Initializing the Population: Random generation of alternate solutions.

• 2. Evaluation: This is done based on the problem’s fitness criterion (objective function) for each candidate solution.

• 3. Evolutionary Process: Selection, crossover, and mutation operations are applied to a population using the standard evolutionary algorithm tradition^24,25; hence, a new generation is produced.

• 4. Local Search: Local search techniques known as memes are used to improve an individual solution. This local search often involves in-depth solution-specific knowledge or heuristics to search the solution space more efficiently.

• 5. End: The algorithm stops when the end criterion has been reached, i.e., a satisfying solution has been reached or the maximum number of iterations has been completed.

These bio-inspired optimization methods can provide opportunities for enhancing the accuracy and speed of deep learning for disease detection applications and for increasing the model interpretability²⁶.

Whale optimization algorithm (WOA)

WOA implements the behavior of trapping social humpback whales’ food (prey). Like other bioinspired algorithms, WOA is based on mimicking whales’ hunting method and particular trapping strategy – the bubble net technique. The whale surrounds the prey and, while doing so, creates special bubbles over the expectant ways. This method of trapping enables for the exploitation. To explore the search for prey by WOA shows the exploration²⁷. As for exploitation, circularity or surrounding diminution enhances the whale’s position. One distinctive feature of WOA is its diverse capability to avoid getting trapped in a local minimum. It simultaneously aims at increasing the merging rate while avoiding convergence to local optima²⁸. Feature selection is a process of selecting a subset of features that will maximize the accuracy of the learning technique and minimize the search space in the learned data. The working mechanism of WOA is shown in Fig. 6.

Fig. 6

Working mechanism of WOA.

The WOA algorithm helps determine the problem space and develop optimal attributes for clinical decisions. The performance was then evaluated using four medical databases, and it was noted that the method could reduce the dimension of medical databases in disease detection with higher accuracy. Because the distribution of classes is uneven, the inter-class difference needs to be increased and the in-class difference reduced; thus, the two-stage WOA introduces a theory of bonds.

A role-based perspective in healthcare innovation

From a role perspective concerning the evaluation of the application of integrative bio-inspired optimization techniques into deep learning frameworks for disease diagnosis, it reflects the effort to define the function of every component involved in the system development and implementation process.

Researcher/developer

The developer helps in the concept, design, and integration of biomimetic mechanisms in the architectural aspects of deep learning and its relation to disease diagnosis. Searches for the biological concepts and the tender algorithm can be used to form the system. This includes the identification and selection of model-sounding ideas as well as the setup of experiments and analysis of the results of the models for disease detection enhancement. Often works with other domain specialists, data analysts, and practitioners to evaluate the developed concepts’ suitability and efficiency.

Data scientist

It is an interesting part of the overall data collection and pre-processing step, as well as manipulating the set for the diseased diagnosing application. That has been described as among the strengths of artificial intelligence due to the possibility of defining datasets required for medical images, electronic health records, genomic sequences, and other data for training or benchmarking²⁹. This involves preparing raw data to one suitable for feeding into deep learning models; this comprises features such as normalization, cleaning, and data preparation. In addition, it identifies features from the feature space of multiple data sources. It encodes prior knowledge of the given domain incorporated into the models and the abilities of metaheuristic algorithms to improve modeling accuracy.

Model architect

Model Architect structures the deep learning design so that it is necessary and possible to identify diseases³⁰. Uses, for example, genetic algorithms, particle swarm optimization, or ant colony optimization to improve the structure, the parameters of the structure, or the procedures used in the training of the model. Introduces new approaches like spiking neural networks or neuromorphic computing, which infers to the biological neurons and improves and develops the models for disease identification.

Evaluator/ethical reviewer

Bio-inspired integration methods are highly adaptive and effective. The following ethical dilemmas of disease diagnostics programs are very crucial. It provides information regarding the number of non-target/endogenous species called accuracy, the ability to identify targets in the presence of sensitivity interferences, and the rest measures during model development using benchmark datasets and its computing cost. Explains privacy, fairness, and transparency as aspects of integrating learning into healthcare solutions.

Clinician/domain expert

Domain experts are prone to endorse disease identification models, which are based on the inputs of specialists and clinicians in designing and testing. Based on clinical relevance, outcome, and marker, it defines the performance and diagnostic criteria for assessing artificial intelligence algorithms. Conducts claims that optimized models are helpful in healthcare networks and satisfy and explainable medical practice.

Regulatory/policy analyst

They ensure that deep learning systems for disease identification comply with the standardized measures, policies, and procedures to the required levels³¹. Assesses the efficiency, safety, and reliability of biomimicry procedures and the implications of such on patients’ well-being and centers. Cooperates with other stakeholders and decision-makers to assess legal, ethical, and social consequences of applying some bio elements to healthcare technologies. In this way, through the connection of actors using the role-based approach, the detection of diseases can be upgraded with the help of integrative bio-inspired optimization in DL systems while overcoming the technical, ethical, and regulatory challenges of healthcare and technological development.

Bioinspired optimization for various disease

Developing disease diagnostic methods based on bio-inspired deep learning optimization can be beneficial for increasing diagnostics in different fields of medicine³². The images are obtained from the Kaggle dataset³³, ISIC dataset⁵⁹, an open-source platform. Figure 7 shows various bioinspired techniques that can be applied to multiple diseases obtained from different datasets.

Fig. 7

Overall flow diagram of bioinspired optimization algorithms.

Breast cancer

Optimizations inspired by biology are incorporated, enabling optimal algorithms for breast cancer detection using mammography and histopathological pictures. Some of the optimization algorithms used include genetic Algorithms (GAs), Particle Swarm Optimization (PSO), and ant colony optimization (ACO) in the optimization of deep learning detection of breast cancer from mammography³⁴ and histopathological images, as shown in Fig. 8.

Fig. 8

Breast cancer images³³.

For example, GAs are useful in generating better neural network architectures that will distinguish between malignant and benign tumors, while PSO optimizes the other parameters of the model that enhance the diagnostic capabilities. For example, ACO helps identify relevant radiomic features from the breast imaging data to find the best diagnosis and therapy plan.

Lung cancer

Some techniques fine-tune deep learning systems for detecting lung nodules and categorizing cancer types from chest X-rays and CTs. In lung cancer detection, the bio-inspired optimization approach enhances deep learning models aiming to classify the chest radiographs and CT scans to identify pulmonary nodules and to categorize tumor subtypes. GAs enhance the over-architectures for nodule detection, whereas PSO helps to tune the model parameters to improve sensitivity and specificity³⁵. ACO helps to pick informative features in data and to determine the stage and prognosis of lung cancer patient images in sample images shown in Fig. 9.

Fig. 9

Lung cancer images³³.

Diabetes

This study attempts to deliver deep-learning models to identify and analyze glycemic control and insulin resistance using electronic health records and wearable sensor information. In the case of diabetes diagnosis and control³⁶, bio-inspiring deep-learning optimization methods are applied to models based on EHR, WS data, and genomics. GAs design and refine neural networks for glycemic control and insulin resistance prognosis, while PSOs tune model aspects for customized glucose tracking. By implementing ACO in feature selection, the risk factors and complications related to diabetes can easily be determined from the extracted features from the various inputs of multimodal data sources, as shown in Fig. 10.

Fig. 10

Diabetes images³³.

Alzheimer’s disease

Currently, optimization methods are used in neuroimaging data from magnetic resonance imaging (MRI) scans and positron emission tomography (PET scans) for early diagnosis of Alzheimer’s and to monitor disease progression. Neuroimaging data such as MRI and PET scans use bio-inspired optimization techniques to detect Alzheimer’s disease at an early stage and progress it’s monitoring. GAs are used to identify the deep learning architecture for automated segmentation of the brain regions and classification of diseases. PSO tunes the model parameters to enhance disease diagnostic quality³⁷. ACO helps identify the most potent imaging biomarkers, which can be used for constructing the risk models from the samples, as shown in Fig. 11.

Fig. 11

Alzheimer’s images³³.

Heart disease

Risk assessment and prognosis of cardiovascular performances have been developed and optimized by deploying deep learning models based on cardiac imaging data and clinical parameters. When it comes to diagnosing heart diseases, even a simple case like applying deep learning optimization techniques in identifying the element that provides the best strength to the model uses Cardiac Imaging data such as echo and angiograms together with clinical factors. GAs design CNN architectures³⁸ for diagnosing cardiac pathologies and estimating the time for cardiovascular incidents, whereas PSO³⁹ tunes the model parameters for developing the risk level. Figure 12 shows the heart images for the input.

Fig. 12

Heart disease images³³.

ACO helps to reduce the dimensionality of cardinal data mining and to find factors that may predict the further course of heart disease and the patient’s response tone.

COVID-19

AI and optimization methods help differentiate COVID-19 pneumonia from other causes based on radiographic chest X-ray and CT images and the prognosis of the disease trajectory. Biological methods are used for the diagnosis and analysis of COVID-19 through radiographs, CT scans, and clinical parameters. GA and PSO⁴⁰ are applied in deep learning models for better recognition of imaging features, as shown in Fig. 13.

Fig. 13

COVID 19 images³³.

COVID-19 pneumonia and better determination of model parameters for assessing severity and allocation of suspected cases. ACO helps in deciding on the variables to be used from the clinical data and laboratory investigations, and hence, it helps in modeling the disease process and the outcome.

Skin cancer

All models are trained and tested for high-accuracy skin lesions and melanoma assessment from dermoscopy and histopathology imagery⁴¹. In cancer diagnosis, especially skin cancer, deep learning optimization through bio-inspired algorithms works in models that utilize dermoscopic images and the histopathological images in Fig. 14.

Fig. 14

Skin tumour images⁵⁹.

From the literature review, GAs are used to design the optimal CNN architecture for melanoma classification and lesion segmentation. At the same time, PSO is applied to fine-tune the model hyper-parameters concerning diagnostics’ efficiency. ACO helps filter out useful dermoscopic features to support the automatic melanoma recognition process with minimal errors.

Stroke

Convolutional mode deep learning networks are trained for the lesion segmentation and stroke subtype prediction task using neuroimaging and clinical features. The neuroimaging data of stroke, such as MRI, CT, and clinical variables, became the target for bio-inspired optimization in diagnosis and prognosis. Figure 15 shows the sample input images of stroke disease.

Fig. 15

Stroke affected images³³.

GAs fine-tune the lesion segmentation and classification of the subtype of stroke. On the other hand, PSO fine-tunes the model’s parameterization for outcome prediction and treatment planning⁴². ACO, having originated from multimodal imaging data, can help with the feature selection process and identify more biomarkers that could be useful in determining the severity of stroke and the patient’s prognosis.

Prostate cancer

Solutions to enhance and analyze multi-parametric MRI are implemented for the accurate detection and localization of prostate cancer lesions. In prostate cancer detection, bio-inspired optimization techniques are used to optimize the model employing the multi-parametric MRI and the histopathological data with the samples shown in Fig. 16.

Fig. 16

Prostate cancer images³³.

GAs design CNN architectures for detection and localization; PSO is employed for the model’s hyperparameters regulating its sensitivity and specificity⁴³. ACO helps choose features from radiometric data that identify and estimate the possibilities of prostate cancer.

HIV/AIDS

Clinical and genomic data are used to predict disease progression, treatment response, and drug resistance using its deep learning model from the images. Heuristic and evolutionary deep learning optimization algorithms work to enhance model performance, achieving HIV/AIDS diagnosis and prognostic accuracy from information sources like clinical data, viral load measurements, and genomic sequences⁴⁴. GAs are used to design unseen architectures for the progression of diseases and reactions to treatments; PSO is used to fit viral load parameters and search for drug-resistant strains. ACO provides support in feature selection to genomic datasets with a choice of genetic markers that characterize HIV/AIDS progression.

Osteoporosis

Implementations are available to start predicting bone mineral density and risk of fractures as much as possible by using the DXA scans and other related clinical risk factors. Optimization In osteoporosis, as shown in Fig. 17, the diagnosis includes bio-inspired optimization of models with DXA scans and clinical markers. GAs generate complex and intricate deep-learning topologies for BMD calculations and fracture risk evaluation; PSO optimizes the achieved model for precision⁴⁵. By extracting feature features from the DXA images and the patient demographics, ACO helps detect osteoporosis early and manage it based on individual patient diagnoses.

Fig. 17

Osteoporosis images³³.

Pancreatic cancer

They involve end-to-end deep learning architectures designed for tumor segmentation and staging based on pancreatic imagery (CT, MRI). In the early screening and diagnosis of pancreatic cancer, bio-inspired deep learning optimization algorithms, such as ANN, CNN, model, and tumor biomarkers, operate over the pancreas’ images, including CT scans and MRI⁴⁶. GAs optimize the CNN architecture for tumor segmentation and grading, while PSO is used to optimize hyperparameters for early diagnosis and treatment models. ACO plays a role in feature selection from radiometric data to determine imaging biomarkers associated with the aggressiveness of the malignancy and survival of the patient with pancreatic cancer in Fig. 18.

Fig. 18

Pancreatic cancer images³³.

Colorectal cancer

The methods of optimization improve the identification and differentiation of colorectal carcinoma and healthy tissues observed in images acquired by colonoscopy and histopathology. The proposed nature-inspired metaheuristic deep learning algorithms are disruptive to the diagnosis of colorectal cancer because they allow for the improvement of distinct diagnostic models influenced by ideas drawn from nature in Fig. 19. Colorectal cancer is one of the common cancers that has been associated with high morbidity and mortality worldwide and has often been challenging to diagnose, especially at its early stage. Nonetheless, new optimization methods that are considered biological evolution, like genetic algorithms (GAs), particle swarm optimization (PSO), and ant colony optimization (ACO), are more effective in dealing with this issue. To improve the deep learning architectures for the detection of colorectal cancer using imaging, GAs optimize network structures specifically for detecting signs of malignant tumors⁴⁷. Thus, while GAs, by repeatedly choosing and integrating architectural elements, the model’s predictive capability of discriminating between malignant and benign tissues.

Fig. 19

Colorectal cancer images³³.

Likewise, PSO adapts characteristics of deep learning models on colorectal cancer data, including the weights and bias of a network, to get the best performance on a classification task. These individual and collective experiences facilitate a reference point of adjusting the parameters so that the model yields more sensitivity to slight changes or differences in histopathological features, making cancer detection more reliable. Furthermore, ACO also helps in screening the characteristic visual patterns from colonoscopy still images and histopathological zones of interest to determine biomarkers linked with colorectal cancer. The ability of ACO to imitate foraging ants and flooding of identified image regions minimizes computation time and increases the model’s interpretability by emphasizing selective prioritization of histological structures or areas of interest. Altogether, these bio-inspired optimization approaches contribute a potentially fruitful direction in the development of colorectal cancer screening, increasing the rate of early detection, enhancing the benefits for patients, and enhancing the efficiency of their treatment.

Hepatitis

Models aim to predict the disease prognosis and how it responds to treatment, with objectives built on clinical information. Artificial neural network learning is an essential subcategory of deep learning optimization in diagnosing hepatitis, as it demonstrates new approaches based on biological processes for refining existing diagnostic models. Hepatitis is an inflammation of the liver⁴⁸, apparent in numerous variants with essentially different origins and symptoms, making its diagnosis and treatment complex. Nonetheless, there are other types of mathematical optimization, such as genetic algorithms (GAs), particle swarm optimization (PSO), and ant colony optimization (ACO), that are bio-inspired and can still be used to solve these problems. Generating accurate predictions of hepatitis patterns can be an unpredictable process; GAs, therefore, allot a critical role in the proper configuration of deep learning structures for the detection of the disorder, as well as for capturing additional diagnostic information on inflammation and severity of the disease. First, while regularly selecting and combining architectural elements, GAs augment the model’s ability to recognize between healthy and damaged liver tissue, thus enhancing the model’s diagnostic performance and reducing misclassification errors.

Likewise, PSO takes the architecture of the deep learning models derived from hepatitis data and adjusts the network weights and/or thresholds of other parameters toward better classification. This dynamic fine-tuning of parameters based on individual and group processes increases the model’s ability to identify and describe the nature of hepatitis further based on the compressed clinical and laboratory findings. Moreover, through the data mining process, ACO helps construct subsets of multidimensional datasets containing clinical features, laboratory results, and measurements of viral loads to determine specific biomarkers and risk factors related to hepatitis disease progression⁴⁹. Following the nature of ants, which search for food sources and apply the procedure of feature ranking to choose the most critical features, ACO improves the model’s interpretability and predictive capabilities for early diagnosis and tailored interventions. Altogether, this bio-inspired optimization presents viable approaches for the development of new diagnostic testing for hepatitis and other forms of liver diseases, which can, in turn, promote early detection and subsequent evaluation of patient experiences and results.

Obesity

Biologically inspired algorithms are becoming significant in solving the complex problem of obesity, which is explicitly acknowledged as one of the devastating health issues in the modern world, defined by the excessive percentage of body fat. These optimization methods learn from natural systems to enable efficient management of obesity and its prevention. For instance, genetic algorithms (GAs) are essential in the genetic profiling and genetic characteristics of the individual who is to be involved in developing a restricted eating plan and exercise regimes, given his or her lifestyle and metabolic requirements⁵⁰. Refining and applying the best combinations of diets and exercises repeatedly, GAs assist the best overweight and metabolic practices to optimize the entire metabolic health process. PSO also plays a part in regulating obesity by supporting the identification of the optimal individual weight loss regime magnitude, type of distribution, necessary calorie intake, and exercise length.

The feature showcased by PSO, where such parameters can be adjusted in cycles to accommodate individual preferences and metabolic rates, further boosts people’s compliance with the plan and increases the chances of sustained weight loss. Moreover, ant colony optimization (ACO) assists in identifying potential optimizing options relating to environmental factors together with social indicators that lead to obesity, like the availability of healthy food, built environment, and social support. Moreover, by emulating the population-driven behaviors of ants and focusing on choosing the most favorable neighborhood conditions for healthy living, ACO contributes to the construction of favorable contexts for developing lasting behavioral patterns and, thus, obesity prevention strategies⁵¹. In general, bio-inspired optimization approaches present new methods for promoting sustainable anti-obesity initiatives while enhancing the capacity of people and populations to make sound decisions and modify weight and health-related behaviors.

Asthma

Forecasting methodology contributes to predicting asthma exacerbations and treatment using clinical parameters and data from respiratory sensors with the images. In asthma detection, biologically inspired computation is our greatest asset since it introduces concepts borrowed from the intricate structure of biological systems to offer far more refined and innovative solutions.

Asthma is a chronic recurrent condition of the respiratory system associated with airway inflammation and increased sensitivity of the smooth muscles of the bronchi. Despite this, there are biologically inspired metaheuristic algorithms⁵², such as gene-based algorithms (GAs), particle swarm algorithms (PSA), and acorn algorithms (ACO), that can be used to solve some of the challenges mentioned. In particular, GAs contribute to fine-tuning the parameters of machine learning models and providing guidance on identifying new features characteristic of the asthmatic population and predicting the severity of the disease. Through the successive adoption and fine-tuning of these model components based on performance feedback, GAs maximize the diagnostic precision and dependability of asthma identification systems.

Comparative analysis

Integrative bio-inspired optimization approaches have demonstrated the potential to improve disease detection in deep learning frameworks⁵³. When these bio-inspired techniques are introduced into deep learning models, the whole system has advantages in both methods, increasing the speed, quality, and stability of disease identification. Different algorithms are taken from various base papers for analysis and the research work is carried out on Matlab 2021a platform. According to the comparative analysis, these integrative approaches are more effective than traditional optimization approaches, which results in more accurate and efficient diagnostic indexes in healthcare systems.

Table 1 Performance comparison on breast cancer using various bio-inspired algorithms.

Particle Swarm Optimization (PSO), Artificial Bee Colony (ABC), Cuckoo Search (CS), Grey Wolf Optimizer (GWO), Dragonfly Algorithm (DA), Genetic Algorithms (GA), Ant Colony Optimization (ACO), Firefly Algorithm (FA), Bat Algorithm (BA), and Whale Optimization Algorithm (WOA) shown in Table 1. Accuracy ranges from 0.87 to 0.93, with Genetic Algorithms achieving the highest at 0.93, indicating the overall correctness of predictions as shown in Fig. 20.

Fig. 20

Breast cancer- performance comparison.

Precision, measuring the proportion of true positives among optimistic predictions, varies from 0.90 (ABC, FA) to 0.95 (GA). Recall (Sensitivity), highlighting the ability to detect actual positive instances, spans from 0.85 (CS) to 0.92 (GA), with GA again leading. F1 Score, a harmonic mean of precision and recall, ranges from 0.87 (CS) to 0.93 (GA), reflecting balanced performance.

Table 2 Performance comparison on lung cancer using various bio-inspired algorithms.

Table 2 compares the performance of various bio-inspired optimization algorithms in lung cancer detection using deep learning systems. Genetic Algorithms (GA) demonstrate the highest accuracy (92.6%) and F1 score (93.5%), indicating their superior capability in optimizing the detection process. Grey Wolf Optimizer (GWO) also performs well, with an accuracy of 91.5% and high precision (92.4%). Ant Colony Optimization (ACO) stands out for its exceptional precision (94.6%) and sensitivity (90.3%), highlighting its ability to identify positive cases, as shown in Fig. 21 correctly.

Fig. 21

Lung cancer- performance comparison.

Despite lower accuracy, the Firefly Algorithm (FA) and Cuckoo Search (CS) show commendable sensitivity, crucial for early disease detection. Overall, GA and GWO are the most effective for lung cancer detection, combining high accuracy, precision, and recall, while ACO’s high sensitivity makes it valuable for reducing false negatives.

Table 3 Performance comparison on diabetes using various bio-inspired algorithms.

Particle Swarm and Whale Optimization Algorithms exhibit high accuracy (88%), precision (85%), and recall (86%), indicating their robustness in fine-tuning models for diabetes prediction, as in Table 3. The Grey Wolf Optimizer (GWO) stands out with the highest accuracy (89%) and similar high performance in other metrics, attributed to its balanced exploration and exploitation capabilities. Artificial Bee Colony and Firefly also demonstrate strong performance, with accuracy at 87% and good precision and recall values. Cuckoo Search (CS) and Bat Algorithm (BA), with an accuracy of 86%, provide solid performance, though slightly lower than PSO and GWO, as shown in Fig. 22.

Fig. 22

Diabetes- performance comparison.

Genetic Algorithms (GA) and Ant Colony Optimization (ACO), both with an accuracy of 85–86%, show promising results, though GA’s performance can vary significantly depending on mutation and crossover rates. While effective, the Dragonfly Algorithm (DA) has a slightly lower performance than GWO and PSO, reflected in its accuracy of 85%.

Table 4 Performance comparison on alzheimer’s disease using various bio-inspired algorithms.

Grey Wolf Optimizer (GWO) stands out with the highest accuracy at 90%, along with superior precision (87%), recall (88%), F1 score (87%), and specificity (92%), as in Table 4. This indicates GWO’s balanced approach to handling the complex patterns associated with Alzheimer’s disease. Particle Swarm Optimization (PSO) and Whale Optimization Algorithm (WOA) also demonstrate strong performance, each with an accuracy of 89%, high precision (86%), recall (87%), and specificity (91%), as shown in Fig. 23.

Fig. 23

Alzheimer’s disease- performance comparison.

These algorithms are practical in fine-tuning models to achieve high predictive performance. ABC and FA show competitive results with an accuracy of 88% and 87%, respectively, indicating their reliability in managing Alzheimer’s disease prediction. Cuckoo Search (CS), with an accuracy of 87%, and Ant Colony Optimization (ACO) and Bat Algorithm (BA), both with an accuracy of 86%, offer solid performance but slightly lower than GWO and PSO. Dragonfly Algorithm (DA) and Genetic Algorithms (GA) exhibit good performance with 86% and 85% accuracy, respectively, though their precision and recall values are slightly lower.

Table 5 Performance comparison on heart disease using various bio-inspired algorithms.

Table 5 above compares various bio-inspired algorithms in the context of heart disease prediction and management. Grey Wolf Optimizer (GWO) stands out with the highest accuracy at 91%, alongside high precision (88%), recall (89%), F1 score (88%), and specificity (93%). This highlights GWO’s efficiency in handling the complexities of heart disease prediction. Particle Swarm Optimization (PSO) and Whale Optimization Algorithm (WOA) also demonstrate excellent performance, each with an accuracy of 90%, high precision (87%), recall (88%), and specificity (92%). These algorithms are adept at fine-tuning predictive models to achieve high overall performance. Artificial Bee Colony (ABC) and Firefly Algorithm (FA) show strong results with accuracy rates of 89% and 88%, respectively, indicating their reliability in heart disease prediction, as shown in Fig. 24.

Fig. 24

Heart disease- performance comparison.

Cuckoo Search (CS), with an accuracy of 88%, along with Ant Colony Optimization (ACO) and Bat Algorithm (BA), both with an accuracy of 87%, offer solid performance though slightly lower than GWO and PSO. Dragonfly Algorithm (DA) and Genetic Algorithms (GA) exhibit good performance with accuracies of 87% and 86%, respectively.

Table 6 Performance comparison on COVID disease using various Bio-inspired algorithms.

Grey Wolf Optimizer (GWO) stands out with the highest accuracy at 92%, along with high precision (89%), recall (90%), F1 score (89%), and specificity (94%). This highlights GWO’s capability to predict COVID-19 cases and manage complex data patterns accurately. Particle Swarm Optimization (PSO) and Whale Optimization Algorithm (WOA) also demonstrate excellent performance, each with an accuracy of 91%, high precision (88%), recall (89%), and specificity (93%) as in Table 6. These algorithms are practical in fine-tuning predictive models to achieve high overall performance. Artificial Bee Colony (ABC) and Firefly Algorithm (FA) show strong results with 90% and 89% accuracy rates, respectively, indicating their reliability in COVID-19 prediction, as shown in Fig. 25.

Fig. 25

Covid- performance comparison.

Cuckoo Search (CS), with an accuracy of 89%, along with Ant Colony Optimization (ACO) and Bat Algorithm (BA), both with an accuracy of 88%, offers solid performance, though slightly lower than GWO and PSO. Dragonfly Algorithm (DA) and Genetic Algorithms (GA) exhibit good performance with 88% and 87% accuracy, respectively. Their precision and recall values are slightly lower, which is typical given their distinct optimization strategies and exploration mechanisms.

Table 7 Performance comparison on skin cancer using various bio-inspired algorithms.

Table 7 above compares various bio-inspired algorithms in the context of skin cancer prediction and management. Grey Wolf Optimizer (GWO) stands out with the highest accuracy at 93%, alongside high precision (90%), recall (91%), F1 score (90%), and specificity (94%). This underscores GWO’s efficiency in accurately predicting skin cancer cases and managing complex data patterns. PSO and WOA also exhibit excellent performance, each with an accuracy of 92%, high precision (89%), recall (90%), and specificity (93%). These algorithms are adept at fine-tuning predictive models to achieve high overall performance. ABC and FA show strong results with 90% and 89% accuracy rates, respectively, indicating their reliability in skin cancer prediction, as shown in Fig. 26.

Fig. 26

Skin cancer- performance comparison.

Cuckoo Search (CS), with an accuracy of 89%, along with Ant Colony Optimization (ACO) and Bat Algorithm (BA), both with an accuracy of 88%, offers solid performance, though slightly lower than GWO and PSO. Dragonfly Algorithm (DA) and Genetic Algorithms (GA) exhibit good performance with 88% and 87% accuracy, respectively. Their precision and recall values are slightly lower, which is typical given their distinct optimization strategies and exploration mechanisms.

Table 8 Performance comparison on stroke detection using various bio-inspired algorithms.

The comparative analysis in Table 8 of bio-inspired algorithms for stroke detection shows that Cuckoo Search (CS) achieves the highest performance with 90% accuracy, 89% precision, 91% recall, 90% F1-score, 88% specificity, and 91.5% sensitivity, indicating its superior effectiveness in identifying stroke cases. PSO and GWO also perform exceptionally well, with PSO showing 89.5% accuracy, 88.5% precision, 90% recall, 89% F1-score, 87.5% specificity, and 90.5% sensitivity, and GWO demonstrating 89% accuracy, 88% precision, 90% recall, 89% F1-score, 87% specificity, and 90.5% sensitivity as shown in Fig. 27.

Fig. 27

Stroke detection- performance Comparison.

Other algorithms like Genetic Algorithms (GA) and Ant Colony Optimization (ACO) show slightly lower performance metrics but are still adequate for specific optimization tasks in stroke detection.

Table 9 Performance comparison on prostate cancer using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for prostate cancer detection in Table 9 shows that Cuckoo Search (CS) achieves the highest performance with 91% accuracy, 90% precision, 92% recall, 91% F1-score, 89% specificity, and 92.5% sensitivity, indicating its superior effectiveness in identifying prostate cancer cases as shown in Fig. 28.

Fig. 28

Prostate cancer- performance comparison.

PSO and GWO also perform exceptionally well, with PSO showing 90.5% accuracy, 89.5% precision, 91% recall, 90% F1-score, 88.5% specificity, and 91.5% sensitivity, and GWO demonstrating 90% accuracy, 89% precision, 91% recall, 90% F1-score, 88% specificity, and 91.5% sensitivity.

Table 10 Performance comparison on HIV/AIDS using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for HIV/AIDS detection in Table 10 shows that Cuckoo Search (CS) achieves the highest performance with 92% accuracy, 91% precision, 93% recall, 92% F1-score, 90% specificity, and 93.5% sensitivity, indicating its superior effectiveness in identifying HIV/AIDS cases. PSO and GWO also perform exceptionally well, with PSO showing 91.5% accuracy, 90.5% precision, 92% recall, 91% F1-score, 89.5% specificity, and 92.5% sensitivity, and GWO demonstrating 91% accuracy, 90% precision, 92% recall, 91% F1-score, 89% specificity, and 92.5% sensitivity as shown in Fig. 29.

Fig. 29

HIV/AIDS- performance comparison.

Other algorithms, such as GA and ACO, show slightly lower performance metrics but are still effective for certain optimization tasks in HIV/AIDS detection.

Table 11 Performance comparison on osteoporosis detection using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for osteoporosis detection in Table 11 shows that Cuckoo Search (CS) achieves the highest performance with 90% accuracy, 89% precision, 91% recall, 90% F1-score, 88% specificity, and 92% sensitivity, indicating its superior effectiveness in identifying osteoporosis cases. Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO) also perform well, with PSO showing 89% accuracy, 88% precision, 90% recall, 89% F1-score, 87% specificity, and 91% sensitivity, and GWO demonstrating 88.5% accuracy, 87.5% precision, 89.5% recall, 88.5% F1-score, 86.5% specificity, and 90.5% sensitivity as shown in Fig. 30.

Fig. 30

Osteoporosis- performance comparison.

Other algorithms, such as GA and ACO, show slightly lower performance metrics but are still effective for certain optimization tasks in osteoporosis detection.

Table 12 Performance comparison on pancreatic cancer using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for pancreatic cancer detection in Table 12 reveals that Cuckoo Search (CS) achieves the highest performance with 89% accuracy, 88% precision, 90% recall, 89% F1-score, 87% specificity, and 90.5% sensitivity, indicating its superior capability in identifying pancreatic cancer cases as shown in Fig. 31.

Fig. 31

Pancreatic cancer- performance comparison.

PSO and GWO also perform well, with PSO showing 88.5% accuracy, 87.5% precision, 89% recall, 88.2% F1-score, 86% specificity, and 89.5% sensitivity, and GWO demonstrating 88% accuracy, 87% precision, 89% recall, 88% F1-score, 85.5% specificity, and 89.5% sensitivity.

Table 13 Performance comparison on colorectal cancer using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for colorectal cancer detection in Table 13 shows that Cuckoo Search (CS) achieves the highest performance with 90% accuracy, 89% precision, 91% recall, 90% F1-score, 87.5% specificity, and 91.5% sensitivity, indicating its superior effectiveness in detecting colorectal cancer.

Fig. 32

Colorectal cancer- performance comparison.

PSO and GWO also perform well, with PSO showing 89.5% accuracy, 88% precision, 90% recall, 89% F1-score, 86.5% specificity, and 90.5% sensitivity, and GWO demonstrating 89% accuracy, 88% precision, 90% recall, 89% F1-score, 86% specificity, and 90% sensitivity as shown in Fig. 32. Other algorithms like Genetic Algorithms (GA) and Ant Colony Optimization (ACO) show slightly lower performance metrics but are still adequate for specific optimization tasks in colorectal cancer detection.

Table 14 Performance comparison on hepatitis using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for hepatitis detection in Table 14 reveals that Cuckoo Search (CS) achieves the highest performance with 89% accuracy, 88% precision, 90% recall, 89% F1-score, 87% specificity, and 91% sensitivity, indicating its superior capability in identifying hepatitis cases as shown in Fig. 33.

Fig. 33

Hepatitis- performance comparison.

PSO and GWO also perform well, with PSO showing 88% accuracy, 87% precision, 89% recall, 88% F1-score, 86% specificity, and 90% sensitivity, and GWO demonstrating 88.5% accuracy, 87.5% precision, 89.5% recall, 88.5% F1-score, 85.5% specificity, and 90.5% sensitivity. Other algorithms, like Genetic Algorithms (GA) and Ant Colony Optimization (ACO), show slightly lower performance metrics but are still adequate for specific optimization tasks in hepatitis detection⁵⁴.

Table 15 Performance comparison on obesity detection using various bio-inspired algorithms.

The comparative analysis of bio-inspired algorithms for obesity detection in Table 15 shows that Cuckoo Search (CS) achieves the highest performance with 92% accuracy, 91% precision, 93% recall, 92% F1-score, 90% specificity, and 93.5% sensitivity, indicating its effectiveness in identifying obesity-related patterns as shown in Fig. 34.

Fig. 34

Obesity- performance comparison.

PSO also performs well, with 91% accuracy, 90% precision, 92% recall, 91% F1 score, 89% specificity, and 92.5% sensitivity. Grey Wolf Optimizer (GWO) and Whale Optimization Algorithm (WOA) also show high-performance metrics, making them suitable for obesity detection tasks⁵⁵. In contrast, Ant Colony Optimization (ACO) shows lower metrics but still provides robust solutions for specific optimization tasks in obesity detection.

Table 16 Performance comparison on asthma disease using various bio-inspired algorithms.

Cuckoo Search (CS) stands out with the highest accuracy (91.5%), recall (92%), and sensitivity (92.5%), indicating its superior performance in asthma prediction. Particle Swarm Optimization (PSO) follows closely with 90% accuracy, 91% recall, and 91.5% sensitivity in Table 16; Fig. 35.

Fig. 35

Asthma prediction- performance comparison.

While algorithms like Ant Colony Optimization (ACO) show lower metrics (accuracy: 85.5%, recall: 86%, sensitivity: 86.5%), they are still adequate for specific optimization tasks. The following Table 17 shows the comparative analysis of different bio-inspired optimization techniques.

Table 17 Comparative analysis of different bio-inspired optimization algorithms.

Challenges and issues

Bio-inspired optimization strategies that can smoothly integrate with deep learning-based disease diagnosis provide the best opportunities. However, various concerns and problems need to be solved to expand the application of these integration frameworks. These issues apply to the computational aspect, the methods used in research, and the practicality of the research concepts.

Computational challenges

High computational cost

Resource intensity: Heuristic methodologies such as gas or PSO, because of their iterative processes and large-dimensional space, generally consume a significant amount of computational power.

Scalability: Amortizing these algorithms to accommodate scalable datasets or sophisticated models poses a challenge, particularly due to scalability issues that call for parallel processing or the use of graphics processing units (GPUs).

Convergence issues

Local Optima: The use of bio-inspired optimization methods is that they are often trapped in the internal extreme points and, as a result, provide less than optimal solutions. Achieving convergence toward the global optimum is a futile task.

Parameter sensitivity: It must be noted, however, that the performance of these algorithms is highly sensitive to one or more parameters, such as the mutation rate for GA and the inertia weight for PSO. These parameters often need to be optimized, which may even be time-consuming.

Methodological challenges

Complexity of model integration

Integration with deep learning: Combining the bio-inspired optimization algorithms with deep learning is not always straightforward because deviation from the original algorithms is often required and often needs to be fine-tuned.

Hybrid models: Developing and validating hybrid models constructed from several bio-artificial composite techniques is essential and challenging. The critical goal is to achieve the maximum of each applied technique’s possibility while minimizing their united disadvantages.

Data quality and quantity

Data requirements: Deep learning models are often data-hungry and need clean data. In medical domains, such data can be scarce, incomplete, or noisy, all characteristics that are undesirable in any learning process.

Feature selection: Bio-inspired methods can effectively detect relevant features within medical data’s extensive attributes.

Practical challenges

Interpretability and explainability

Black box nature: When deep learning is employed and bio-generated techniques are used to fine-tune the model, there is usually a lack of tractability, which means it is challenging to determine the rationale behind the model’s choices.

Clinical trust: The lack of interactivity can become an issue when accepting such models in clinical practice, although it is crucial for clinical practice.

Validation and generalization

Over fitting: It was also mentioned that while bio-inspired optimization can be helpful, it can lead to over fitting if applied with fine-tuned models to the training data⁵⁶. Cross-validation proved to be a very effective method for boosting models’ performance when dealing with new data.

Validation standards: Setting sound validation criteria and approaches to evaluate the effectiveness of such fine-tuned models in different care scenarios is challenging.

Regulatory and ethical issues

Compliance: One of the big challenges is to ensure that bio-inspired deep learning systems follow the rules and regulations required by bodies such as the Food and Drug Administration (FDA).

Ethical Considerations: Below are some of the areas of concern when it comes to addressing policy considerations in the use of AI in healthcare: ethical issues such as the right to privacy and especially managing data on patients, the aspect of informed consent in the use of the models, and the problem of prejudice in the predictions made by the AI.

Integration into clinical workflow

Usability: For the models to be adopted, special consideration should be given to developing interfaces that are friendly to the end user and how simplified the integration into the clinical setting is.

Training and education: Successful use of these new diagnostic tools in a given clinical setting means that more time and resources are needed to properly train healthcare professionals.

Research and development challenges

Algorithm development

Innovation: The never-ending cycle of designing new biologically inspired algorithms and refining existing algorithms for better performance forms a continuous research challenge.

Benchmarking: this study, the authors provided reliable findings and compared various bio-inspired optimization techniques for the detection of diseases, which is essential to constructing a platform to compare the results in future research⁵⁷.

Cross-disciplinary collaboration

Interdisciplinary knowledge: multidisciplinary framework is needed for applying such techniques to detect diseases involving biology, computer science, and medical sciences.

Collaboration: deed, the processes of developing recommendations for clinical practice, designing new studies, and analyzing the results clearly require the complex cooperation of doctors, scientists, and methodologists.

Altogether, integrative, bio-inspired optimization methods for disease diagnosis in deep learning systems have high potential. However, the identified computational, methodological, and practical tasks must be solved to introduce them into clinical practice⁵⁸ successfully. By far, efforts to continue researching, strengthening interprofessional collaborations, and creating effective frameworks and guidelines will play an essential role in addressing these challenges and unlocking the potential of these sophisticated diagnostic tools.

Research directions

Deep learning in medicine applied to disease diagnosis promises a radical shift in Medical diagnostics through integrative bio-inspired optimization. As datasets’ size increases and the modeled features become more affluent, it is necessary to turn to methods based on the Princess of natural systems. Some significant trends emerging from this field of study are as follows, each focusing on unique studies and analyses.

Advanced Bio-inspired algorithms and hybrid approaches

One potential line of research is the improvement of the current state-of-the-art metaheuristics and creating new metaheuristic algorithms that take bio-inspired principles from both natural systems and technologies. This method combines two or more bio-inspired algorithms, including GAs, PSO, and ACO, where the integration of several of these enhances the model’s performance. For example, one can boost the application of GA, which has a high ability to search ample solution space, and the other can benefit PSO, which performs well in the fine-tuning stage to obtain a higher optimized technique. Metaheuristic algorithms can also be flexible and self-adjustable, the so-called ‘adaptive metaheuristics,’ where some or all of the parameters of the algorithm are adjusted during the optimization process to enhance the convergence rates and the quality of the solutions. These advancements focus on developing innovative deep-learning models designated explicitly for the intended function of disease identification.

Scalability, efficiency, and interpretability

Computational issues are also among the research priorities since they help address complex problems that address computational requirements. The choice made in this paper to employ both parallel and distributed versions of bio-inspired algorithms is to take advantage of current and robust HPC infrastructure for handling large datasets and sophisticated models involved in medical diagnosis. Another area that needs to be addressed is the efforts to increase the rates and position of algorithms, such as better initialization methods and improved mutations and crossover operations. However, certain limitations have been observed; as the complexity of deep learning models increases, the interpretability and explainability of such models are practically difficult to maintain. Pursuing more straightforward interpretation approaches, such as post hoc attribution methods for applying AI and better visualization tools, would enable clinicians to grasp how models make decisions, making the process more trustworthy and giving a better route to achieving regulatory approval.

Multimodal data integration and personalized medicine

Research into integrating and further optimizing models that utilize multimodal data is still in its early stages. Imaging findings, genomic data, and clinical information can reveal a disease’s essential details and help enhance diagnostic performance. A lot of effort has been made to develop techniques for integrating data so that the data collected from various sources can become more coordinated to improve disease diagnosis.

These research directions demonstrate the tremendous potential of bioinspired optimization methods in increasing the performance of deep-learning networks for disease diagnosis. Thus, by countering the computational, methodological, ethical, and practical issues, the researchers will obtain better diagnostic tools that are accurate, efficient, and clinically relevant. Further development in this area will create a fundamental understanding of disease diagnosis and patients’ treatment, allowing for more advanced, reliable, and efficient healthcare strategies.

Conclusion and futurescope

Bio-inspired optimization methods have been effectively utilized in improving deep learning systems for health diagnosis, which provides new paradigms to the toughest issues prevailing in healthcare systems today. Incorporating principles from natural systems such as genetic algorithms, Particle swarm optimization, and Ant colony optimization can enhance the Deep learning model’s performance, accuracy, and efficiency. These techniques can help create models that are maximally specialized and resistant to errors when identifying and differentiating further diseases, starting with cancer and ending with cardiovascular diseases, relying on various medical data. However, despite making distinct progress and inventions, several issues are yet to be addressed. The high computation costs and the problem of convergence, as well as the necessity of using large, high-quality datasets, are some of the challenges. However, another aspect that plays a significant role in applying these complicated models is the possibility of interpretation. These models are explainable, and clinicians can easily understand their decision-making process, which is crucial for gaining their trust and, by extension, that of the regulators. More research and development have to be conducted to overcome these challenges, and the following aspects should remain in focus: better algorithms, hybrid optimization methods, and better tools needed for clarifying models.

Future scope

The prospect of further developing different integrative bio-inspired optimization techniques that can be used in disease detection is quite bright, and several directions could be explored in further research. One such area is improving the metaheuristic algorithms, which use features of several bio-inspired optimization methods to obtain higher efficiency. These hybrid methods can be further subdivided in that they can involve dynamic tweaking of respective parameters, types of attacks, and detection mechanisms according to the needs of the various medical applications. Also, the improvement of parallel and distributed computing will be helpful for these optimization processes and make technology for vast medical data analysis possible. Another important direction is the integration of multimodal data sources. Future studies may direct the efforts toward improving the techniques for integrating the imaging, genomic, and electronic health records data to build even more accurate diagnostic models. These changes can drive the delivery of personalized and definite healthcare services, enhancing the results. In addition, the advance of explainable AI techniques will facilitate the softening of the gap between the efficiency of the optimization models and the practicability in clinical practice so that clinical decision-makers can trust the recommendations made by those AI systems.

Finally, managing the ethical and regulatory stance will be essential in deploying these technologies in healthcare. Future research should aim to produce guidelines and measures to reduce biases and protect the patient’s confidentiality. Substantial clinical implementation and experimentation will be required to prove if these optimized models are safe and effective, thus creating a foundation for their application in regular clinical practice. The upcoming years will require research cooperation with medical departments and regulatory bodies to address such potential challenges and make the opportunities for Bio-inspired optimization techniques more prominent in the diagnosis process. Thus, despite the difficulties in applying integrative bio-inspired optimization techniques for disease detection and diagnosis in deep learning systems, the opportunities for enhancing the healthcare system cannot be overemphasized. For additional studies and advancements in the diagnostic models, which will be more accurate, efficient, and easier to interpret, more work is essential in this area, which will enhance disease diagnosis and the patient’s healthcare conditions all over the world.