The traditional Indian system of medicine promotes overall health and wellness through personalized therapies and a detoxifying process. Therapeutic emesis (TE), or vamana karma, is one of the bio-pentavalent purification procedures used to treat deranged kapha ailments that include metabolic, respiratory, and dermatological conditions such as psoriasis and eczema [1]. TE is conducted in 3 phases (ie, preparatory phase, main phase, and postpurificatory phase). In the main phase, volunteers are subjected to induced vomiting through the provision of emetic herbal medicines. After initiation of the process, the individual expels the vitiated content, concurrently evident through physical changes (concurrently evident include the onset of nausea, increased salivation, abdominal discomfort, sweating, piloerection, lacrimation, and the progressive expulsion of gastric contents, reflecting effective mobilization and elimination of morbid bio-humor) [2]. This procedure is carried out early in the morning and takes approximately 1 hour, although it sometimes takes more than an hour to complete.

Previously, TE has been assessed subjectively by trained physicians through visual inspection–based emesis indicators, such as patient gestures that include nausea, discomfort, and distress and vomitus characteristics such as color, consistency, volume, and odor. Nevertheless, dependence on human knowledge may introduce observer bias and interindividual variability, affecting reliability in clinical outputs [3]. Furthermore, extended therapy is often time-consuming and poses challenges for physicians in observing and documenting continuously, potentially creating inconsistency and reducing standardization. Studies suggest that human error is one of the most significant contributors to decreased productivity and quality. Multifactorial interaction among stress, repeated tasks, fatigue, and workplace circumstances impairs cognitive performance, attention, and decision-making accuracy [4]. Similarly, the variable rapid nature of emesis creates difficulties in manual tracking due to the rapid onset of the process, thus limiting real-time monitoring.

Research highlights that the integration of artificial intelligence (AI) techniques in image analysis for clinical decision support showcases their potential in diverse areas. Today’s health care practices have changed significantly with the help of AI and machine learning (ML), which help with more accurate diagnoses, custom interventions, and informed decisions [5,6]. Previous studies have substantiated that the use of AI-assisted imaging yields superior accuracy and specificity compared to traditional methods [7]. Further cutting-edge developments using YOLOv9, TensorFlow, and residual neural network (ResNet) models have yielded newer insights in scope for such assessment methods.

There have been many studies conducted on vamana karma focusing on its efficacy in diseases and biophysical parameters [8-10]. Presently, no study has been proposed to digitally analyze the procedure. This study has been designed to integrate for the first time an AI component for TE analysis. The primary objective is to support physicians in accurate interpretations of purification type and guiding patients to an appropriate postpurificatory diet. The primary aim is to assess appropriate signs and symptoms of TE through content analysis of vomitus by developing and validating an AI model using digital image processing and an AI algorithm. Additional objectives focus on developing a dataset of TE for training and validating an AI model, training the TE model for real-time assessment, and validating the performance of the developed AI model by comparing its assessments with those of expert Ayurvedic physicians.

In the traditional system of medicine, TE is a key intervention for body detoxification, which requires documentation of all evident changes during the process. Recent studies have underscored the adoption of the YOLOv9 approach for detecting abnormal crowd behavior to promote crowd safety [11]. Correspondingly, another study used the same methodology in traffic hazards for public safety using the GC-YOLOv9 algorithm [12]. Additionally, there are studies in which image processing was used for microscopic sputum analysis and the velocity distance support algorithm was used for sputum monitoring to detect phlegm stagnation [13,14]. Contemporarily, one study was conducted to automate emesis in Suncus murinus using a convolutional neural network and a self-attention mechanism [15]. Such studies draw attention to developing a model that can automate emetic episodes in traditional sciences. Moreover, the convergence of such approaches remains unexplored. Leveraging existing technologies could enable real-time monitoring for automation, enhance reliability, and improve patient outcomes.

This is a prospective, observational, single-center study aimed at developing and validating an AI-based framework for the objective assessment of vamana karma (TE), a classical purification therapy in Ayurveda. The study adheres to the Standards for Reporting Diagnostic Accuracy–Artificial Intelligence reporting guidelines [16]. It will ensure data handling with model development, validation, and performance. A checklist has been provided in Multimedia Appendix 1.

The study will be conducted at a tertiary Ayurveda research and teaching institute with an established panchakarma (bio-purification methods) center. A dedicated clinical setup will be designed to ensure uniform data collection across participants. The study workflow is depicted in Figure 1A and Figure 1B.

To achieve uniformity in data collection, we will adopt the following guidelines throughout the process:

After providing informed consent, adult participants aged between 18 and 65 years will be enrolled in the study. Inclusion and exclusion criteria will be strictly adhered to for participant safety and data reliability. Inclusion criteria are adults clinically recommended for TE as per classic Ayurvedic guidelines for kapha-dominant conditions such as asthma, skin diseases, and metabolic disorders. Participants must be willing to consent to video recording during the procedure. Exclusion criteria include individuals with contraindications for vamana karma, including pregnancy, lactation, cardiac conditions, anemia, or known hypersensitivity to emetic medicine. Individuals unable or unwilling to provide informed consent will also be excluded.

A total of 40 high-resolution videos (approximately 70 minutes each) were used, yielding over 10,000 frames annotated for vomit detection. This dataset size was considered adequate for training a deep learning model given the high temporal density of labeled events and the pretrained backbone used (YOLOv9 and ResNet). A preliminary learning curve analysis indicated performance saturation beyond this dataset size. Thus, for this study, we targeted 50 participants.

The complete process will be conducted in 3 phases, as shown in Figure 2. Phase 1, also known as the preparatory phase (purva karma), starts with the administration of deepana (digestive stimulants) and pachana (digestive agents), followed by snehapana (internal oleation) for 5 to 7 days accompanied by abhyanga (massage) and svedana (sudation) for 1 day. In this 1-day period, a kapha-aggravating diet will be administered. Thereafter, the second or main phase (pradhana karma) starts on the vamana karma day with abhyanga (massage) and svedana (sudation). Participants will be allowed to take the following emetic formulations early in the morning: (1) rice gruel (consistency, color, and texture), (2) milk (2 L), (3) 8 g of medicine (madanphala powder; Randia dumetorum Linn.)+5 g of madhuyashti powder (Glycyrrhiza glabra Linn.)+2 g of vacha powder (Acorus calamus Linn.)+5 g of salt+honey (q.s.), (4) 30 g of madhuyashti kwath (decoction of G glabra Linn.) powder dissolved in 1 L of hot water, and (5) salt water (20 g of salt dissolved in 1 L of water; 1 glass=380 mL). The number, content, and force of vomitus bouts will be recorded. Vomit will be collected in calibrated jars for visual analysis and recorded via video.

In the third or postprocedure phase (paschat karma), after completion of vamana karma, depending on the type of purification, the participants will be gradually shifted to samsarjana karma (postdetoxification dietary regimen) to normalize the weak digestive fire.

Video data will be captured throughout the procedure to record vomitus events, facial expressions, and physical gestures. The data will form the primary source for model training and validation.

Video data will be processed through frame extraction and manual annotation to create labeled datasets for ML. More than 11,000 frames will be annotated for vomitus event detection, whereas approximately 700 to 800 cropped images will be classified as per classic parameters. All these annotations will serve as the ground truth for supervised learning and will be validated by experienced Ayurveda physicians to ensure accuracy. Data labeling details are listed in Table 1.

Descriptive statistics will summarize participant demographics and baseline characteristics. ML model performance will be reported using accuracy, precision, recall, F₁-score, and area under the receiver operating characteristic curve.

Model validation will use 5-fold cross-validation techniques to ensure robustness. Agreement between AI outputs and physician assessments will be evaluated using the Fleiss κ with standard interpretation thresholds of <0.40 for poor agreement, 0.40 to 0.75 for fair to good agreement, and 0.75 for excellent agreement.

All analyses will be conducted using Python libraries (scikit-learn [Google Summer of Code] and TensorFlow [Google Brain Team]) and the SPSS software (IBM Corp) for statistical analysis.

After training both the YOLOv9 and ResNet models, testing will be conducted in the following structured way. The dataset will be divided into 70% for training, 15% for validation, and 15% for testing. The test set will consist of unseen video samples containing vomiting and nonvomiting events to ensure unbiased evaluation. The YOLOv9 model will be tested on the test dataset to detect vomit regions frame by frame, output bounding boxes, and confidence scores. To test the complete pipeline, the full video will be passed through YOLOv9 followed by ResNet. The pipeline will be evaluated on frame-level accuracy (correct vomit identification across frames), event-level accuracy (correct identification of full vomiting events), and processing time per frame (frames per second) to measure real-time feasibility.

K-fold cross-validation (eg, K=5) will be performed to verify model robustness and reduce overfitting. Additionally, to prevent data leakage, participant-level splitting will be used during model evaluation. All frames from a given participant will be kept within a single fold to ensure that no individual’s data appear in both the training and test sets. This ensures realistic generalization performance.

To enhance clinical interpretability, gradient-weighted class activation mapping visualization will be explored on the ResNet classifier outputs to highlight the image regions contributing to vomit type classification. This step supports explainability and builds clinician trust in the AI’s decisions.

The primary outcome is accuracy (or F₁-score) of vomit event detection and classification at the participant level. Secondary outcomes are inference time per frame and classification consistency across participants.

Ethics approval for the study was obtained from the Institutional Ethics Committee of All India Institute of Ayurveda, New Delhi, India before initiation (325/19.12.2022/PhD-04/2022 [reissued in 2023]). All participants will provide written informed consent before recording after being informed about the study objectives, the voluntary nature of participation, and their right to withdraw at any time without consequences. Videos will be anonymized by removing identifiable facial features and stored in encrypted format (Advanced Encryption Standard–256) on a secure local server with access restricted to authorized researchers.

While the potential of AI integration in Ayurveda is significant, it also raises ethical and cultural challenges. AI-based evaluation of deeply traditional and personalized procedures may be viewed with skepticism by practitioners who value individualized clinical judgment. Hence, AI tools should be positioned as decision support systems rather than replacements for practitioner expertise. Transparent algorithms, patient consent, and data privacy must remain central to any implementation.

The study seeks to test a novel way to assess TE by using an AI model. The proposed framework will evaluate the outcomes more accurately and ease for the physician by automating the vomiting event. Once validated, this framework could serve as a model for the application of AI in other traditional medicine procedures, thereby contributing to the broader field of integrative and digital health. Previous studies in medical imaging and diagnostic AI have established the capacity of ML algorithms to identify subtle patterns in biomedical data—such as radiographs, endoscopic images, and pathology slides—that often escape human perception. In traditional systems of medicine, similar AI-based models have been used for pulse diagnosis, tongue image interpretation, and facial feature analysis, suggesting the need for deployment [17-21]. However, no prior work has attempted this in the domain of vomitus analysis, leaving a critical research gap that needs to be addressed. The framework can be generalized to other panchakarma procedures or Ayurvedic practices involving visual inspection provided that adequate domain-specific annotations are available. Its modular design allows for retraining with site-specific data for broader clinical adoption. The study establishes the foundation for image-based pattern recognition models that could one day assist practitioners in evaluating the shuddhi lakshana (signs of purification) objectively.

The study will help enhance diagnostic precision and documentation. Moreover, it enables longitudinal data collection that may reveal new correlations among patient profiles, therapy outcomes, and procedural parameters, which were otherwise difficult to quantify manually.

The study’s shortcomings include a small sample size, possible observer bias in vomitus labeling, and limited generalizability due to single-center data. Data will be collected in a controlled clinical setting, limiting real-world diversity. TE is a seasonal therapy traditionally administered during specific periods of the year, which naturally limits the availability of eligible participants and, consequently, the sample size for ML model development and validation. Furthermore, the digital interpretability of Ayurvedic features such as pitta (bio humor responsible for metabolic activities) and aushadhi (medicine) remains challenging. Despite these limitations, the study is an important step toward developing a validated framework for combining AI with traditional assessments. More comprehensive multicentric research is necessary to validate the proposed concept across diverse populations and practice settings.

Future work will focus on validating the framework across larger, more diverse patient populations and exploring its application in real-time clinical decision-making. Although this work focuses on model development and internal validation, future studies will involve clinical validation across multiple panchakarma centers to assess model robustness under real-world variations in lighting, camera positioning, and practitioner styles.

This work presents one of the first attempts to apply deep learning for objective analysis of the vamanakarma process in Ayurveda. By combining YOLOv9 for vomit detection and ResNet for classification, the framework achieves promising accuracy in automated vomit identification. The results will demonstrate the potential of AI-assisted analysis in traditional medicine, although further clinical validation and expansion across multiple centers will be necessary before deployment in real-world settings.