Diseases and disorders disrupt one or more biological processes. Despite differing mechanisms, they can produce similar symptoms or physiological responses. Clinicians diagnose and develop effective treatment strategies using observations from multiple sources and scales, ranging from a macroscopic view of a patient’s overall physical and mental health to microscopic sequencing or cell imaging data. While integrating different types of datasets has been an active area in multi-modal machine learning in biomedicine for several years, fusing data across different scales remains less common.

As Steyaert et al. discussed in a Perspective last year, combining diverse, multi-modal data that span space and time can overcome the limitations of single data modalities, which may be noisy or incomplete. Fusing data from various scales provides a promising strategy for designing models that are sensitive and robust enough to diagnose, prognose and assign treatments1. Such approaches mimic the holistic philosophy used by clinicians of incorporating specific patient information from diverse sources. For instance, in this issue, Endo et al. outline a data-driven, multi-modal and multi-scale approach for sub-typing Parkinson’s disease (PD)2.

Motivated by the link between motion impairment and brain function in PD, a neurodegenerative disease, Endo et al. integrate video-captured motion data of participants with functional magnetic resonance imaging (fMRI) data2. A main challenge in integrating two incongruent data types is effective preprocessing. Endo et al. address this by distilling key biomarkers from the two data streams, reducing their dimensionality2. From the motion data, they estimate the change in joint angles during locomotion, and from the fMRI data, they isolate the functional connections between different parts of the brain. These processed features of the original data have similar graph-like form and can be combined to generate a low-dimensional representation. Thereafter, participant data are clustered into sub-categories from the two multi-scale data streams. This approach allowed the authors to accurately classify three subtypes of PD and propose digitally derived biomarkers that signify them.

Other recent developments in machine learning-based modelling for neurodegenerative disease categorization follow a similar multi-modal, multi-scale tenet. In a recent Article in Nature Medicine, Xue et al. focus on classifying different types of dementia3. This is important for effective treatment planning and identifying the aetiology of each type, but it is challenging with predictive models based on a single data type, due to overlapping clinical presentations. Mimicking clinical reasoning, the authors developed a model that combines a wide array of data, including brain imaging, demographic and neurological data, from over 50,000 people, to categorize them into various stages of cognitive decline.

A challenge in advancing multi-scale machine learning, as is often the case, is data availability. Useable multi-scale biological data pertaining to individual participants’ diseases are valuable but rare. Aside from biomedical data acquisition’s being taxing and time consuming, there are physiological effects to consider. Aspects of the body can change rapidly in response to diseases, and if data acquisition occurs over weeks or months, one individual data stream, such as single-cell RNA-sequencing data, acquired from a patient at a specific timepoint, may not adequately represent the patient when an additional data stream, such as microscopy imaging, is acquired at a different timepoint. Organization is therefore required for realistic, reliable and useful conclusions to be drawn from multi-scale biomedical data. The use of retrospective data is unlikely to give new insights in this direction, as many openly available and widely used data sources are not multi-modal. There are existing examples of careful, ethical and organized multi-modal data sharing — for instance, the UK Biobank, MIMIC-IV4 and The Human Connectome Project5 databases — but these are generalized rather than targeted to specific diseases and disorders, and so are unlikely to offer clear insights into them. New high-quality datasets that are multi-scale and inspired by specific clinical applications and challenges, such as those presented by Endo et al.2, Xue et al.3 and others, will bolster progress in this direction.

Biological systems operate through complex processes that work harmoniously at multiple scales. As showcased by Endo et al.2, careful selection of biomedical data, taking into account biomarkers at different scales, can lead to predictive models capable of categorizing diseases accurately, potentially enabling more-targeted treatment strategies.