Artificial Intelligence Applications in Ayurveda: A Systematic Review

Abstract

Artificial intelligence is applied to Ayurveda in a growing literature, but no systematic review has mapped what that literature contains, which computational methods it uses, or how reliably it reports them, and existing reviews of AI in traditional medicine treat Ayurveda alongside Traditional Chinese Medicine and other systems, where it is a small minority of the evidence. This review searched Scopus, PubMed, Web of Science and Google Scholar from database inception to August 2026 for studies applying AI or machine-learning methods to Ayurveda or Indian traditional medicine. Records were deduplicated with a four-pass deterministic cascade (DOI, PMID, normalized title, title prefix) followed by manual review of multi-signal fuzzy-matched pairs. Eligibility required an explicit AI or machine-learning method (C1), Ayurveda or Indian traditional medicine as the study domain (C2), and at least one of five predefined application domains (C3). Studies whose methodological contribution was recognition of plant, leaf or species from images were excluded, including those that added a therapeutic lookup to the recognized species. Every included study was read in full and coded on ten reporting items, and the 14 studies that predict an individual’s disease or constitution were assessed for risk of bias with PROBAST+AI. Screening, extraction and coding were carried out by a single reviewer, the protocol was not registered, and the review is reported according to PRISMA 2020. Of 1,253 records, 635 remained after screening and deduplication, and 42 studies met all criteria (38 from database searching, 4 from hand-searching); 38 of the 42 appeared from 2024 onward. Disease or symptom prediction with therapeutic linkage was the most common domain (22 studies), and drug discovery, phytochemistry and molecular modelling were the least common (8). Scopus returned all 38 database-traced studies, and no other database retrieved the included study that Scopus had missed. Image-based plant recognition accounted for 156 of 597 excluded records, more than any other exclusion reason. Full-text assessment and re-checking removed seventeen studies that had passed title and abstract screening: eight with no learning method, four reviews, three in which Ayurveda was not the study domain and two that reported no results, one of the last carrying undisclosed machine-generated text alongside unverifiable references. Eight of the 42 studies showed none of the ten reporting defects and fifteen showed four or more. Molecular-modelling studies carried fewer defects than the rest (mean 1.12 against 2.91, two-sided Mann-Whitney p = 0.020), but seven of the eight were journal articles, and the difference could not be separated from publication venue (adjusted rate ratio 0.63, 95% CI 0.29–1.38). All 14 prediction-model studies were at high risk of bias, most often because of their data sources and analysis. AI work on Ayurveda is concentrated in symptom-to-remedy recommendation systems, often built on datasets whose size is not reported, and is thinly represented in compound-level discovery; much of it is published in venues not indexed by the major bibliographic databases, and fourteen screened records retrieved as AI research contained no learning component.