• Use of AI in Ayurveda has benefits, limitations and challenges. Success of AI depends on collaboration of Machine Learning developers with Ayurvedic physicians and latter’s ability to familiarize themselves in technology. 

Introduction

Artificial intelligence (AI), simply defined as the ability of a machine to mimic or surpass human intelligence, is capturing the attention of experts in technology, business, medicine and even philosophy. By using modern computing technologies to analyze millions of data points, computer programs with AI solve complex problems previously considered beyond human intellectual capability.

One sub-field of AI with substantial implications in healthcare is machine learning (ML) which concerns techniques that enable algorithms to learn to perform specific tasks like classifying large volumes of data or predicting outcomes based on historical information. ML algorithms are used to create AI models for specific purposes like predicting a patient’s response to an intervention. In Western medicine, AI models are generated through co-creation by physicians and computer scientists. Scientists write the code that constitutes the model and train the model using datasets provided by physicians. Multiple iterations and evaluations are conducted before a model is deemed ready to use. An AI model is as robust as the data it is trained on.

AI is entering Ayurveda as well, albeit more slowly and in a fragmented manner. Our review of extant scholarship over the past 5 years indicates that Ayurvedic scholars propose the deployment of ML algorithms on repeatable and predictable aspects like identifying medicinal plants and automating the detection of VPK doshas and subsequently, individual prakriti.

Pattern generation is efficient in such cases, however, Ayurveda distinguishes itself from Western medicine by its focus on holistic wellbeing and psychosomatic healthcare. There is usually not a linear connection among the multitude of attributes affecting holistic wellbeing, limiting the application of AI in such nuanced cases. For example, chronic stress is a mainstay among today’s youth and is exacerbated by their lifestyle choices, coping skills and cognitive abilities. While AI does not yet offer a platter of ready solutions in such cases, we found studies indicating its potential application in psychosomatic health. 

For example, (1) demonstrated that ML models could accurately predict stress levels, focus, long-term engagement and relaxation levels of an individual using various input parameters including demographics, brainwaves and lifestyle factors. Studies of this nature make it easier for holistic practitioners to suggest practical tips to minimize stress and eliminate lifestyle diseases. This is a promising application of AI in Ayurveda.

This article was first published in the Aryavaidyan Journal October to December 2022 issue.

Extant research on the applications of AI on Ayurveda is summarized below.

Medicinal Leaf Discovery

Ayurvedic medicinal plants are not rare and are often accessible to the public. However, identifying them requires high levels of experience and expertise. Various attributes of the plant including its bark, color, texture, and leaf structure are used by experts to recognize useful herbs. This time-consuming manual effort is now being automated by efficient plant recognition systems deploying ML algorithms.

For example, (2) compared multiple ML models to determine plant species from the images of their leaves. Using a training dataset containing 64000 images, they tested multiple ML models for accuracy and found that convolutional neural networks could identify plant species with over 97% accuracy. Similar results have been found by (3) whose ML model trained from an Ayur Leaf dataset containing leaf samples from 40 medicinal plants in Kerala, classified them with 97% accuracy.

One interesting feature about ML advancements in Ayurveda is that it eliminates the need for expensive hardware. A low-budget intelligent system using Raspberry Pi 3 (a one-piece computer the size of a credit card) and an associated camera was used by (4) to recognize details in images of medicinal leaves. Traditional textual analytics methods involve an approach called Bag of Words (BoW) where the model identifies and counts word combinations to accurately determine key themes and sentiments of a conversation.

Similarly, ML techniques using images adopt a visual BoW approach. By building a training dataset containing clear sample images, the authors determined that ML models could achieve an accuracy of 98% in plant recognition, a number substantially higher and more consistent than manual inspection of leaf texture and shape. Having a low-cost portable intelligent plant recognition system is a substantial leap in Ayurvedic medicine as it eliminates the need for constant availability of experts and decreases the possibility of errors in species identification.

Automating the Determination of VPK doshas

Intelligent systems are being used to predict VPK doshas with over 90% accuracy. Pulse recognition using nadispandanam is one method used in Ayurvedic diagnosis. ML algorithms using techniques such as decision trees, LASSO regression, Elastic Net and random forest clustering are being deployed in complement with pressure and optical sensors to capture pulse waves akin to the Ayurvedic doctor’s spandanam (5,6).

For example, (7) used a hybrid method involving artificial neural networks and fuzzy inference logic to predict the doshas and prakriti of an individual with 92% accuracy even with a small training dataset. Ensemble techniques using ML algorithms that integrate multiple models to incrementally refine parameters and enhance predictive accuracy are also in use; the AdBoost algorithm has been demonstrated to offer more accurate predictions than traditional methods (8).

Visual recognition (darsanam) is also being automated through facial recognition using a variety of ML techniques. These models operate in three phases: first they detect faces and their boundaries from possibly cluttered backgrounds, then identify relevant features like skin, hair, eyes and face shape, and finally, categorize these appropriately to determine the individual’s prakriti. ML algorithms that extract and segment facial features use parametric classifiers like Gaussian clusters.

For example, skin color maybe detected using a combination of multilayer perceptrons (neural network models that perform approximations) and Gaussian classifiers (9) or statistical ML approaches like Random Forest. Similarly, Support Vector Machine (SVM) is an ML algorithm used to determine hair and non-hair regions in an image. SVM maps a multidimensional space, like a human head, into data points in a way they may be easily categorized. In doing so, it creates a separator between categories, and transforms data points such that a plane of separation may be easily identified. Areas of the image corresponding to the eyes may be extracted and input into the ML model. By categorizing features in this manner into the VPK types, the algorithm determines the individual’s prakriti by adding up the score for each type. The final outcome is the type with the majority score. 

Challenges Implementing AI in Ayurveda

Our modern lifestyle complicates Ayurvedic diagnosis by causing molecular changes that may not be directly aligned with the doshas documented in the seven traditional Ayurvedic phenotypes. The stratified approach to treating disorders based on predominant dosha imbalances lends itself to data-driven analyses and pattern detection using algorithms. However, the unprecedented intrapersonal genetic or epigenetic variability arising from contemporary lifestyle choices including geography, exposure to chemicals or hazardous plastics at work, consumption of highly processed foods, and ethnicity of parents complicates matters. Practices like upaasayam facilitate periodic monitoring of patient response and frequent changes in medication to accelerate recovery. This trial-and-error method, however, raises challenges for ML-based drug prediction.

For one, the complex interplay among multiple factors makes diagnosis less straightforward. Second, it introduces subjectivity around the doctors’ expertise. ML models are prediction machines based on pattern generation and recognition. Commercial ML models can process millions of data points arising from thousands of input variables, however its accuracy is based entirely on its training data. 

When clear correlations between input and output are unavailable, or when the training data is missing adequate data points on diagnosis of such patients, the algorithm may identify spurious correlations based on available data, resulting in unusable and potentially dangerous outcomes. 

Similarly, current AI interventions in medicinal plant identification rely entirely on visual identification, noise removal and extraction of relevant features to accurately categorize leaves. While this is helpful, it omits other key aspects of leaf identification including rasa (taste), guna (qualities), veerya (potency- ushna/seeta) and prabhava (specific action), resulting in a less-than-holistic approach.

Accurate drug prediction from ML algorithms will require clear definition of potential input variables and detailed documentation of the patient’s recovery at each stage to be used as training data. ML-based analysis and determination of phenotypes is currently nascent and our review indicates that more scholarly focus is necessary to capture the nuances of treatment especially for chronic diseases that are on the rise.(1) is a step in the right direction, however, more scholarly exploration is warranted.

Despite the advancements in ML, AI-guided diagnosis carries risks and must be exercised with caution. ML algorithms are incredibly complex, and require advanced coding skills to generate, making them prone to errors. Incorrect coding in algorithms may lead to false negatives in the norm of missed diagnosis, unnecessary treatments due to false positives, or more generally incorrect interventions.

A specific challenge for Ayurveda is the lack of access to large datasets incorporating sufficient quality and in terms of comprehensiveness and treatment variation, which may be used to train ML models. Insufficiently-trained models may generate spurious correlations as they are impacted by the noise in input data. Incorrect classifications may also result when the statistical distribution of data used in practice shift from what was used in training; for example, a model trained purely on an Indian genetic dataset may operating on a patient with dissimilar genetic composition, like a European.

To mitigate these issues, we suggest that the role of AI be clearly defined – algorithms must be assistive tools to the physician and not autonomous tools that replace them. Physicians must continue as primary drivers of the patient workflow and be present in all stages of diagnosis to avid errors and potential harm to the patient. Next, physicians and governing bodies must define procedures and protocols to evaluate and approve ML models, especially when they are being generalized across multiple populations.

ML algorithms are prone to misuse by physicians because they are created by computer scientists and with little physician input. This makes it the physicians’ job to learn how the algorithm operates, a herculean task leading to confusion, frustration and eventual avoidance of the tool by the physician. Extant training programs in Ayurvedic medicine do not equip physicians with the knowledge of AI, thereby limiting the possibilities of informed decision making through the tool and increasing the possibility of error. 

To this end, we suggest that healthcare professionals sensitize themselves to the emerging possibilities of AI and ensure that they are closely aligned with the design and development of ML algorithms in their organization. We also suggest that educational sessions on AI in healthcare and specially Ayurveda be developed and disseminated widely within the medical fraternity.

Finally, ML algorithms always carry the possibility of perpetuating current systemic inequities in healthcare due to factors like gender, age and income. Gender-based discrimination is rampant in healthcare; (10) demonstrates how female patients are invisibilized when reporting pain. Women are also systemically diagnosed later than men for the vast majority of diseases (11). Such bias, when allowed to seep into the input data used to train the ML model, leads to perpetuation of these biases. Bias may also be geographic; models missing global representativeness of training data are likely to be regionally biased and generate incorrect diagnosis on patients from another geography.

The global debate around algorithmic fairness has awakened AI developers and users alike; global movements like explainable AI are first steps in this direction. We believe there is much ground to cover before AI achieves complete fairness and transparency.

Therefore we suggest that ML developers collaborate with Ayurvedic physicians to inspect training data and ensure that it is balanced and representative around key attributes of potential discrimination.

Conclusion

AI proffers substantial potential for diagnosis and treatment of patients using Ayurvedic principles. Collecting sufficient and reliable data around epidemiology, clinical investigation, prognosis and course of treatment is a primary concern. Using AI requires a mindset shift among physicians to share necessary information without compromising patient privacy, along with an upgrade in their technical skills to upload medical documents online. Broadening of physicians’ skill sets is also a need, as interdisciplinary expertise is necessary to competently deploy big data for Ayurvedic treatment. The power of AI lies in discerning patterns in complex systems in a manner that transcends human capability; however, AI can never replace the judgment and emotional intelligence of doctors as they comfort and encourage patients through recovery; this is after all, why healthcare exists.

Author Bio - Anjana Karumathil is Associate Professor of Practice at the Indian Institute of Management, Kozhikode. Dr. Ramesh Varier is Chief (Clinical Research) at Arya Vaidya Sala, Kottakkal.

Article available online/offline on: Aryavaidyan Journal.

No part of this article may be reproduced in full or part without written permission of the Aryavaidyan Journal.

To read all articles on Ayurveda

References

1. Morande S. Enhancing psychosomatic health using artificial intelligence-based treatment protocol: A data science-driven approach. International Journal of Information Management Data Insights. 2022 ;2(2):100124.

2. Paulson A, Ravishankar S. AI Based indigenous medicinal plant identification. In:2020 Advanced Computing and Communication Technologies for High Performance Applications (ACCTHPA) 2020 Jul 2(pp. 57-63). IEEE.

3. Dileep MR, Pournami PN. AyurLeaf: a deep learning approach for classification of medicinal plants. In: TENCON 2019-2019 IEEE Region 10 Conference (TENCON) 2019 Oct 17 (pp. 321-325). IEEE.

4. Shailendra R, Jayapalan A, Velayutham S, Baladhandapani A, Srivastava A, Kumar Gupta S, Kumar M. An IoT and machine learning based intelligent system for the classification of therapeutic plants. Neural Processing Letters. 2022;54(5):4465-93.

5. Tiwari P, Kutum R, Sethi T, Shrivastava A, Girase B, Aggarwal S, Patil R, Agarwal D, Gautam P, Agrawal A, Dash D. Recapitulation of Ayurveda constitution types by machine learning of phenotypic traits. PLoS One. 2017;12(10):e0185380.

6. Pogadadanda H, Shankar US, Jansi KR. Disease diagnosis using ayurvedic pulse and treatment recommendation engine. In: 2021 7^th International Conference on Advanced Computing and Communication Systems (ICACCS) 2021 Mar 19 (Vol. 1, pp. 1254-8). IEEE.

7. Madaan V, Gayal A. An Adaptive Neuro Fuzzy Inference System for Predicting Ayurvedic Dosha. In: 2019 4^thInternational Conference on Information Systems and Computer Networks (ISCON) 2019 Nov 21 (pp. 335-9). IEEE.

8. Rajasekar V, Krishnamoorthi S, Saracevic M, Pepic D, Zajmovic M, Zogic H. Ensemble Machine -learningmethods to predict human body constituencies. Computer Science. 2022;23(1):117-32.

9. Phung SL, Bouzerdoum A, Chai D. Skin segmentation using color pixel classification: Analysis and comparison. IEEE Transactions on Pattern Analysis and Machine Intelligence. 2005;27(1):148-54.

10. Samulowitz A, Gremyr I, Eriksson E, Hensing G. “Brave men” and “emotional women”: A theory-guided literature review on gender bias in health care and gendered norms towards patients with chronic pain. Pain Research and Management. 2018.

11. Westergaard D, Moseley P, Sørup FK, Baldi P, Brunak S. Population-wide analysis of differences in disease progression patterns in men and women. Nature Communications. 2019;10(1):666.