Can an algorithm learn part of what an experienced vaidya observes during Prakriti assessment? Peer-reviewed studies now show that machine-learning models can classify carefully selected Vata-, Pitta-, and Kapha-dominant research participants from structured phenotypic data. That finding is scientifically interesting, but it does not mean that a computer has reproduced the full process of Ayurvedic diagnosis.
The strongest published results come from small, highly curated cohorts of people with clearly dominant constitutions. The models were trained using features and labels supplied through Ayurvedic assessment, so they should be understood as tools for testing reproducibility and reducing large questionnaires—not as independent proof of every classical claim or as replacements for clinical judgement.
What Classical Prakriti Assessment Includes
In Ayurveda, Prakriti refers to an individual’s constitutional pattern. Charaka Samhita, Vimana Sthana 8.95, relates its formation to factors including the characteristics of the reproductive elements, the condition of the uterus and season, maternal diet and conduct, and the interaction of the mahabhutas. The passage recognises Vata-, Pitta- and Kapha-dominant constitutions, combinations of doshas, and a balanced constitution.
Prakriti assessment is not based on a single sign. In Vimana Sthana 8.94, Charaka instructs the physician to examine constitution together with present morbidity, tissue excellence, compactness, bodily measurements, habituation, mental strength, digestive and food capacity, exercise capacity, and age. Classical clinical reasoning therefore places constitution within a broader examination rather than treating a questionnaire score as a complete diagnosis.
Modern questionnaires commonly translate these observations into items concerning build, skin, hair, appetite, digestion, temperature tolerance, sleep, activity, memory, speech and emotional tendencies. Seven broad categories are usually used: Vata, Pitta, Kapha, Vata-Pitta, Pitta-Kapha, Vata-Kapha and a relatively balanced type. The distinction between Prakriti and Vikriti is essential: Prakriti concerns constitutional tendency, whereas Vikriti concerns the person’s present deviation or disorder.
Why Standardisation Became a Research Priority
Traditional assessment depends on training, observation and interpretation, so different instruments or practitioners may not always classify the same person identically. Standardisation is therefore important when researchers wish to compare genomics, metabolism, treatment response or disease patterns across sites.
A 2025 critical review in Frontiers in Medicine identified 64 distinct Prakriti assessment tools used in 94 studies published between 1987 and 2024. Only 20 tools had undergone any form of validation, and none satisfied all nine criteria in the review’s validation framework. The CCRAS-Prakriti Assessment Scale and the Ayurveda Child Personality Inventory performed best in that evaluation, each meeting seven of the nine criteria. This review shows both the scale of current interest and the continuing lack of a universally accepted gold standard.
What the Verified Machine-Learning Studies Found
| Study | Data and Method | Verified Result | Important Limitation |
|---|---|---|---|
| Prasher et al., 2008 | Biochemical measurements and whole-genome expression in strongly defined constitutional groups | Reported differences in biochemical variables and gene-expression categories among extreme Vata, Pitta and Kapha groups | This was biological-correlation research, not an AI diagnostic-accuracy trial |
| Govindaraj et al., 2015 | Genome-wide analysis of 262 well-classified men selected after screening 3,416 participants | Reported constitution-associated genetic signals, including a highlighted association involving PGM1 | Male-only, highly selected extreme phenotypes; results do not establish a universal genetic test |
| Tiwari et al., 2017 | 133 phenotypic features; LASSO, elastic-net and random-forest models; 147-person discovery cohort and 96-person external cohort | In external validation, class sensitivity ranged from 79.3% to 100%, depending on model and constitution, with reported specificity above 90% | Participants represented extreme Vata, Pitta or Kapha types rather than the full range of mixed constitutions |
| Khatua et al., 2023 | Dense neural-network analysis of 233 extreme-Prakriti records assembled from two regional cohorts | Showed that deep learning could reduce the number of phenotypic variables while retaining useful cross-cohort classification performance | The data remained small, selected and dependent on pre-existing expert labels |
| Venkatesh et al., 2025 | Critical review of 64 assessment tools used in 94 studies | Found considerable methodological variation and incomplete validation across available instruments | Most tools are not yet validated across diverse populations or routine clinical settings |
How to Interpret the Reported Accuracy
The most frequently cited machine-learning evidence is the 2017 PLOS ONE study by Tiwari and colleagues. The investigators analysed 147 people with strongly expressed Vata, Pitta or Kapha phenotypes from a western Indian cohort and tested the models in 96 independently assessed participants from northern India. External-validation sensitivity varied by method and class: LASSO achieved 93.1% for Kapha, 82.7% for Pitta and 94.7% for Vata; elastic net achieved 96.5%, 86.2% and 97.3%; and random forest achieved 100%, 79.3% and 97.37%, respectively.
These figures should not be converted into the blanket statement that AI can diagnose every person’s Prakriti with more than 90% accuracy. The research deliberately selected clear single-dosha phenotypes, while many people assessed in practice have dual-dosha or less sharply differentiated constitutions. The labels also came from Ayurvedic phenotyping; the algorithm learned patterns within that framework rather than discovering the categories without prior Ayurvedic input.
The study’s internal holdout result also requires context. A 100% result was obtained in a test subset of only 16 people. Small test sets can produce unstable estimates, so the larger external-validation results are more informative. Future models need prospective testing in broader age groups, sexes, regions, ancestry groups and mixed constitutional types.
Genomic Findings: Association Rather Than Final Proof
Biological studies have reported differences among carefully selected constitutional groups, but they do not show that Prakriti is determined by one gene or that a consumer DNA test can establish constitution. The 2008 study by Prasher and colleagues reported group differences in biochemical measurements and whole-genome expression patterns. The authors presented the work as evidence that extreme constitutional phenotypes could be investigated biologically.
In 2015, Govindaraj and colleagues screened 3,416 people and analysed 262 well-classified male participants. Their genome-wide study reported 52 single-nucleotide polymorphisms at the study’s stated significance threshold and highlighted a relationship between Pitta Prakriti and variation involving PGM1, a gene connected with glucose metabolism. The carefully selected sample improved contrast between groups, but it also limits generalisation to women, mixed constitutions and unselected populations.
These studies support continued investigation of Prakriti as a form of phenotypic stratification. They do not justify deterministic claims such as “Vata has neurological genes,” “Pitta has inflammatory genes,” or “Kapha has fat-storage genes” as fixed rules for every individual. Pharmacogenomic applications remain a research hypothesis requiring replicated, clinically relevant studies.
The Microbiome Evidence Is Exploratory
Research has also examined whether gut and oral microbial communities differ among Prakriti groups. A 2018 study of 113 healthy people from a relatively homogeneous rural western Indian population found a broadly shared gut-microbiome structure along with differential enrichment of some microbial taxa among Vata-, Pitta- and Kapha-dominant participants.
A later study involving 272 healthy participants reported a shared microbial core and some constitution-associated signatures in oral and gut samples. These are association studies. They do not establish that microbiome data alone can reliably diagnose Prakriti, that doshas directly create particular bacteria, or that changing the microbiome changes a person’s constitution.
Digital Pulse and Image Analysis Remain Early-Stage
Instrumented pulse recording is a legitimate area of biomedical engineering, but it should not be confused with complete validation of Ayurvedic pulse diagnosis. A 2007 publication described Nadi Tarangini as a system for acquiring and digitising radial-pulse waveforms with pressure sensors. Subsequent pilot work explored measurable vascular characteristics, including arterial stiffness.
Those publications demonstrate that pulse waves can be recorded and analysed objectively. They do not establish fixed autonomic profiles for every Vata, Pitta and Kapha individual.
Facial-image analysis, smartphone photography and other computer-vision approaches are also being proposed. The 2025 review found that such approaches remain methodological frameworks rather than generally validated clinical instruments. Performance reported from a development dataset should not be treated as evidence of safety, fairness or accuracy in the public.
Reasonable Uses of AI in Ayurveda
AI may be useful for reducing lengthy research questionnaires, checking scoring consistency, identifying which features contribute most to a classification, and enabling larger observational studies. Official digital resources also exist: the Ministry of Ayush hosts Ayusoft, and the CCRAS Prakriti Assessment Scale has an associated web portal and training manual.
These resources do not make every online “dosha quiz” clinically valid. A responsible tool should disclose its questionnaire, target population, training data, reference assessment, validation method and error rates for each constitutional class. It should also distinguish Prakriti from current symptoms instead of interpreting temporary digestive, sleep or emotional changes as permanent constitution.
AI systems may reproduce regional, sex-related or assessor-related biases present in their training data. They may also express uncertain results with unjustified confidence. Validation in one selected Indian cohort does not guarantee equivalent performance in another region, country or clinical population.
What AI Cannot Replace
Charaka Samhita, Sutra Sthana 9.6, describes four essential qualities of a physician: sound theoretical knowledge, extensive practical experience, dexterity and purity. A classification model may assist with organising observations, but it does not possess clinical experience, ethical responsibility or the capacity to examine the whole patient.
Ayurvedic care also requires consideration of present disease, strength, digestion, habituation, age, season, medicines, contraindications and changes over time. A constitutional label alone cannot determine treatment. The verified literature does not support replacing consultation with an automated score or using AI output to prescribe herbs, purification procedures or restrictive diets without professional review.
The most defensible conclusion is that machine learning can reproduce aspects of expert-labelled Prakriti classification under controlled research conditions, especially for strongly dominant Vata, Pitta and Kapha phenotypes. It has not yet validated every classical attribute, solved disagreement between assessment methods, or demonstrated reliable diagnosis across all seven constitutional categories.
Disclaimer: This article reviews classical and scientific literature and does not constitute medical advice. For clinical Prakriti assessment, diagnosis or treatment, consult a qualified Ayurvedic physician or other appropriate healthcare provider. Do not begin herbs, supplements, restrictive diets or therapeutic procedures solely on the basis of an app or AI-generated classification.
References
- Charaka Samhita — Vimana Sthana 8.91-95
- Charaka Samhita — Sutra Sthana 9.6-10
- Frontiersin (frontiersin.org)
- Whole genome expression and biochemical correlates of extreme constitutional types defined in Ayurveda (2008), PubMed
- Nature (nature.com)
- Journals (journals.plos.org)
- Classification of Ayurveda constitution types: a deep learning approach (2023)
- Western Indian Rural Gut Microbial Diversity in Extreme Prakriti Endo-Phenotypes Reveals Signature Microbes (2018), PubMed
- Exploring the signature gut and oral microbiome in individuals of specific Ayurveda prakriti (2021), PubMed
- Nadi Tarangini: a pulse based diagnostic system (2007), PubMed
- Significance of arterial stiffness in Tridosha analysis: A pilot study (2017), PubMed
- Ayusoft (ayusoft.ayush.gov.in)
- Prakriti (prakriti.ayush.gov.in)
- CCRAS
The IIT Bombay and CSIR research on Prakriti classification is work I have followed since 2018. The 72-94% accuracy range is significant variance and the paper-by-paper analysis shows why. The high-accuracy results use multimodal data including genomic markers. Phenotypic assessment alone lands in the 70-80 percent range. Still impressive for a categorical classification.
Useful post on AI and Ayurveda. I would still ask a practitioner before changing medicines.
Started the prakriti protocol my Ayurvedic doctor recommended last summer and my energy has been noticeably more consistent. This explains the mechanism.
I have taken 3 online Prakriti quizzes and gotten 3 different results. The consistency problem is exactly what the machine learning research is trying to solve. If a trained algorithm using validated questions is more consistent than informal quizzes, that is useful even if it cannot replace the experienced practitioner.
Useful post on AI and Ayurveda. I appreciate that it does not oversell the result.
As a data scientist, the study designs described here are actually reasonable for what they are trying to do. The confusion matrix approach to evaluating classification accuracy is appropriate. My concern would be the ground truth problem, how do you validate that the machine has correctly identified Prakriti when the gold standard is practitioner judgment, which itself has interrater reliability issues?
The section on AI-assisted Prakriti assessment apps being developed for direct consumer use raises regulatory and safety questions that are not addressed here. If the app suggests wrong Prakriti and the person follows an inappropriate protocol for their constitution, who is responsible for the harm?
How well do these machine learning models handle mixed dosha constitutions when applied outside the selected extreme groups
The democratization argument for AI Prakriti assessment is important for making Ayurvedic care accessible in areas where trained vaidyas are not available. Not everyone has access to a practitioner who can spend 45 minutes doing a classical assessment. A validated algorithmic tool would not replace that but would make something better than nothing available.
The study’s reliance on highly curated cohorts makes me wonder if the results will hold up in a more diverse clinical setting
The IIT Bombay Prakriti research has been on my radar for a while. What I find genuinely interesting is the challenge of encoding something like pulse diagnosis — where a vaidya is reading multiple layers simultaneously — into feature vectors that a model can actually use. Did the paper discuss how they handled the subjectivity between different practitioners during the training data collection? That seems like the biggest methodological hurdle.
It’s interesting to see algorithms used just to check consistency of scoring rather than to replace the vaidya’s judgment
Reducing lengthy questionnaires could save time for researchers studying genotype constitution links
I’m cautious about interpreting any single gene association as proof that a dosha is genetically fixed
After reading I plan to look at the CCRAS Prakriti Assessment Scale portal to see how the labels are generated
The AI-assisted Prakriti apps already on the market are deeply inconsistent and several have been criticized by Ayurvedic scholars for oversimplifying the assessment. Celebrating the machine learning validation without acknowledging the quality of existing implementations is misleading.
For someone whose Prakriti assessment changes between online questionnaires and a trained practitioner’s assessment, which should be trusted for making health decisions? The machine learning accuracy range of 72-94% means 6-28% of people are getting incorrect classification. That is clinically significant.
The consent and privacy implications of collecting the kind of multimodal physiological and genomic data required for high-accuracy Prakriti classification are significant. Who owns that data and how it might be used commercially is a question the article does not address.
The idea of using ML to standardize Prakriti assessment is genuinely exciting. The inter-rater variability among Ayurvedic practitioners has been a long-standing limitation.
The training data for most Prakriti ML models is biased toward educated urban populations who tend to seek Ayurvedic assessment. Generalizability across socioeconomic groups is unproven.
The part about AI and Ayurveda feels realistic. The article avoids making it sound like a quick fix.