Machine Learning and the Limits of Classical Ayurvedic Food Data: Evidence from Caraka Samhita for Model Evaluation and Diet Assessment

Abstract

Machine-learning studies increasingly draw on classical Ayurvedic texts, but how data derived from such texts should be tested has received little attention. We coded 368 food entries from Caraka Samhita Sutrasthana 27 into a knowledge base of Rasapanchaka attributes and Dosha effects and compared a classifier that uses food names alone with classifiers that use Rasapanchaka attributes to predict the Dosha effects the text records. We then applied the knowledge base to twelve months of mess menus from four hostels of Jawaharlal Nehru University (336 day-menus), assessed the Prakriti of 197 resident students with a standardised scale, and recorded clinically assessed Vikriti in eight residents of single-Dosha constitution. Two sources of inflated accuracy were found. A classifier that used only food names scored 100% on its training data but no better than the majority class on unseen foods (Vata 51.1%, equal to the majority share). Because Caraka often describes several foods in one passage, 81 of the 156 entries used for attribute models were exact copies of another entry, and allowing copies to fall in different cross-validation folds raised macro-F1 to 0.58 to 0.72. With copies kept together, Rasapanchaka attributes predicted the text’s Dosha effects moderately for Vata (macro-F1 0.49 to 0.50; AUC 0.71), no better than the classical taste rule for Pitta (0.40 to 0.44 against 0.42) and at chance for Kapha (AUC 0.51 to 0.54). Scored through the knowledge base, all four hostel diets came out as net Dosha-decreasing, mainly because 72.0% of Caraka’s directional statements describe a Dosha being lowered and most foods carry no statement on Pitta; recoding four staple grains by their recorded taste reversed the mean Kapha score of three hostels from -18% to +47%. Of the 197 residents, 177 (89.8%) were of dual-Dosha (Dvandvaja) constitution and 20 (10.2%) of single-Dosha (Ekadoshaja) constitution, and Prakriti did not differ by sex. Vata appeared in the Vikriti of six of the eight residents examined, although every hostel diet was predicted to lower Vata. Models built on classical food texts should be tested on held-out textual descriptions rather than single entries, and diet scores derived from such texts describe what the text chooses to mention until they are validated against measured outcomes.