Ayur.AI, Chennai, Tamil Nadu, India | December 2022
Evidence based personalized ayurveda is critical to realize quantitative metric driven Ayurveda diagnostics and therapeutics which can lead to large scale deployment and adoption of Ayurveda globally. The relative distribution of three doshas (prakriti) VATA, PITTA, KAPHA determines the unique phenotype of a person. Any imbalance in these doshas leads to vikriti. Currently, prakriti and vikriti determination requires deep expertise and is qualitative. The purpose of this study is to quantitatively determine the prakriti and vikriti by machine Learning and deep learning techniques which can act as a digital assist to Ayurveda doctors and patients.
For prakriti prediction, a conversational AI based app is developed which captures patients’ phenotypic characteristics including physical, anatomical, physiological and psychological parameters. Machine learning methods such as Random Forest (RF), K-Nearest Neighbor (KNN), Support Vector Machine (SVM) and Naive Bayes (NB) are used for assessing prakriti based on captured features. To determine vikriti (dosha imbalance), patient’s finger pulse is captured using a smartphone camera. A deep neural network based digital nadi pariksha technique was developed which uses captured blood pulse waveform as an input to assess the vikriti. The reference labels for both prakriti and vikriti were given by an Ayurveda doctor.
The prakriti prediction model is trained with 74 different phenotypic characteristics, obtained using a conversational AI based questionnaire, on a cohort of 505 individuals. The dataset is divided into 70% training set and 30% testing set and the machine learning models are trained using RF, KNN, SVM and NB. To improve the accuracy further, we have also explored an ensemble of RF and SVM methods. The results show that the optimized Random Forest model gives a training accuracy of x% and a validation accuracy of 68% for mixed prakriti prediction. Similarly, the vikriti prediction model trained using a custom designed deep neural network with a ReLu activation function showed a training accuracy of x% and a testing accuracy of 58%.
The present study is a first of its kind study demonstrating the utility of machine learning and deep learning techniques for prakriti and vikriti prediction. In future, using more patient’s data and advanced machine learning and deep learning models, we aim to increase the prakriti and vikriti prediction accuracy to >90%. The app incorporating the developed algorithms can be used for rapid prakriti and vikriti prediction and hence can be used as a digital assist in Ayurveda practice. An Ayurveda based digital wellness app is planned that can provide accurate diagnosis using the developed algorithms as well as personalized diet, herbal supplement, yoga and meditation recommendations.