Table 1 Overview of common AI/ML algorithms
From: Opportunities and challenges of artificial intelligence in hepatology
Algorithmic model | Category | Data type | Brief definition | Strengths | Limitations | |||
|---|---|---|---|---|---|---|---|---|
Images | Numerical | Sequence ¹ | ||||||
Linear regression | Machine Learning (ML) | ✓ | Predicts a continuous outcome by fitting a linear relationship between input features (e.g., predictor variables) and the target. | Simple, fast, and interpretable; works well on small and complete datasets. | Assumes linearity and independence, sensitive to outliers and unsuitable for complex non-linear relationships. | |||
Logistic regression | ✓ | Estimates the probability of a binary outcome (e.g., Yes/No) by applying a logistic transformation to inputs features. | Interpretable and efficient for binary classification, fast to compute, useful as a baseline model. | May underperform for complex non-linear or high-dimensional feature interactions. | ||||
Random forest (RF) | ✓ | Ensemble of multiple decision trees where each tree brings in a vote and the ensemble “forest” created averages output results, improving accuracy and reducing overfitting. | Robust to background noise and overfitting, handles high-dimensional and mixed data effectively. | Less interpretable, slower on large datasets, may still overfit if not tuned. | ||||
Support vector machine (SVM) | ✓ | Classifies data by constructing an optimal hyperplane (i.e., subspace of one dimension less than the input space) in a multi-dimensional space. | Effective for complex classification tasks, can handle high-dimensional and non-linear data. | Requires careful parameter tuning, kernel choice (i.e., evaluation of the data’s linearity and complexity) matters. Computationally intensive on large datasets. | ||||
LASSO regression | ✓ | Linear regression method that adds a penalty to coefficients to shrink less informative features to zero, performing both regularisation and feature selection at the same time. | Performs feature selection and regularisation simultaneously, enhances model simplicity, interpretability, and generalisation while reducing overfitting. | Can only capture linear relationships, could exclude relevant correlated features. | ||||
Extreme gradient boosting (XGBoost) | ✓ | Optimised implementation of gradient boosting (i.e., ensemble method of sequential weak-learners where each corrects the residual errors of the prior) using decision-trees, with added regularisation for speed and scalability. | High predictive accuracy, can handle missing values and non-linear relationships, easily scalable for large datasets. | Higher risk of overfitting and less interpretable, requires careful hyper-parameter tuning. | ||||
Neural networks (NNs) | Feedforward NN | Deep Learning (DL) | ✓ | Multilayer architecture where information moves one-way through layers to capture complex non-linear relationships. | Can learn complex, non-linear relationships between features; adaptable across data types. | Performance is heavily dependent on the availability of large, high-quality datasets to allow generalisation. Internal “black box” structure limits interpretability. Model design requires careful optimisation of architecture, layer depth, and learning parameters, as shallow networks may underperform while overly complex one’s risk overfitting. Computationally demanding and hardware-intensive training, which can hinder scalability and real-time clinical deployment. | ||
Convolutional (CNN) | ✓ | Deep architecture specialised in processing grid-like data (e.g., images) by using convolutional filters to automatically detect and extract features (e.g., edges, shape, patterns). | Excellent for image segmentation and classification with automated feature extraction, high accuracy in visual tasks. | |||||
Artificial (ANN) | ✓ | ✓ | General multilayer neural architecture that processes numerical or mixed data through multiple non-linear transformations for pattern recognition. | Versatile across numerical and categorical data, can capture non-linear relationships. | ||||
Transformer | ✓ | ✓ | ✓ | Attention-based architecture that models contextual relationships across sequential inputs (especially effective for speech and text). | Can capture long-range dependencies in sequential data; strong performance in text, speech, and multimodal inputs. | |||
Large language models (LLMs) | DL/Gen AI | ✓ | ✓ | ✓ | Large-scale Transformer trained on extensive multimodal datasets, capable of reasoning with language. | Broad task adaptability (e.g., generating, summarising, extracting), scalable across modalities. | Prone to factual errors (“hallucinations”), less medical-specific trained due to being computationally costly, regulatory/ethical/privacy issues. | |
