Answers are grounded in this paper and cite their sources.
Multimorbidity—the concurrent presence of two or more chronic long-term conditions—affects more than 37% of the global adult population and consumes 55%–80% of healthcare expenditure in high-income systems, yet artificial intelligence (AI) diagnostic systems remain overwhelmingly disease-specific, creating a profound structural mismatch with the multi-disease complexity characterizing the majority of real-world clinical encounters across twelve review-defined disease groups mapped to World Health Organization (WHO) ICD-11 chapters. This systematic review, following PRISMA 2020 and registered with OSF.IO (Registration DOI number: 10.17605/OSF.IO/3HZ45; Date of Registration: 25 October 2026), evaluated advanced AI technologies for simultaneous multi-domain disease detection and prediction. Nine electronic databases, PubMed/MEDLINE, IEEE Xplore, Scopus, Web of Science, ScienceDirect, Springer Link, Wiley, MDPI, and Google Scholar, were searched for peer-reviewed literature published from January 2015 to 31 August 2026. Database-specific Boolean searches (full strings in Supplementary Table S1 ) combined AI terms with terms for the twelve review-defined disease groups. Quality appraisal was done using QUADAS-2 and PROBAST tools (Cohen's κ = 0.84 refers to study screening agreement). Of 4,215 records identified (3,368 after removing 847 duplicates; 612 full texts assessed; 332 excluded), 280 studies met the eligibility criteria and were included in the qualitative synthesis, and 185 provided quantitative performance data for a descriptive (non-pooled) synthesis. Deep learning models such as CNN, LSTM, ViT, and GNN reported high but heterogeneous discrimination across the twelve disease groups (AUC: 0.85–0.991). These values come from different tasks, datasets, and validation designs; they are not pooled estimates and do not show superiority of any architecture. Three of the highest-reporting single-disease studies (retrospective and not directly comparable) include: diabetic retinopathy detection (AUC 0.991; n = 128,175); cancer whole-slide image classification (AUC 0.98; n = 44,732); and cardiac arrhythmia detection (AUC 0.97; n = 91,232 ECGs). Convolutional neural networks were the most popular network type (26% of studies), while the main programming language was Python (68%). Three repeated gaps can be seen in the field of research in general, and those are: 1) lack of representation of people from developing countries in the training data; 2) scarce evidence from randomized prospective trials showing improved patient outcomes; and 3) an underdeveloped regulatory pathway for multi-disease AI as Software as a Medical Device (SaMD). Realizing simultaneous multi-disease AI at clinical scale demands five coordinated priorities: standardized cross-category benchmarks, demographically inclusive datasets with mandatory low- and middle-income country representation, prospective randomized trial validation, federated privacy-enhancing infrastructure designed to support GDPR, HIPAA, and DPDPA 2023 compliance (federated learning alone does not establish compliance), and internationally harmonized regulatory governance for multi-disease Software as a Medical Device authorization. Limitations of the review itself (English-only restriction, heterogeneity, overlapping public datasets, publication bias) are discussed in Section 8.5 . Systematic Review Registration: https://doi.org/10.17605/OSF.IO/3HZ45
No takes yet. Share an insight, caveat, or question.
Choudhary et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: