Appendicitis Prediction Tool Logo
×

Dharma: Appendicitis Model

Dharma is a novel, non-linear, clinically grounded and interpretable machine-learning based clinical scoring system for diagnosis and prognosis of acute appendicitis in patients presenting with acute abdominal pain. It utilizes routinely available clinical parameters and point-of-care ultrasound findings to provide diagnostic and prognostic risk estimates.
Thus providing an evidence-based clinical decision support.

Dharma Framework:
The framework consists of three key components: 1. Dharma_Imputer : Intelligent imputation of commonly missing features (CRP, urinary ketones, free-fluids) using decision trees pre-trained on appendicitis data. It also handles missing appendix diameter feature as an independent signal, mimicing real-world clinical reasoning.
2. Dharma_Dianostic Model : Balanced random-forest model.
3. Dharma_Severity Model : Highly sensitive random-forest model for severity screening for confirmed appendicitis cases by the upstream diagnostic model.

SHAP Explanations:
The model uses SHAP (SHapley Additive exPlanations) values to interpret predictions. Each SHAP value quantifies the contribution of a specific input feature to the model’s output, relative to a base value. The base value represents the model’s expected output, and the sum of the SHAP values shows how much each feature pushes the prediction above or below this baseline.

Built by a clinician, for the clinicians.

LITERATURE (peer-reviewed): PLOS Digital Health
EXPLORE Dharma: GitHub Repository Link

PERFORMANCE METRICS:
Diagnostic Performance (cross-validation)
AUC-ROC: 0.93-0.99
Sensitivity: 90%-95%
Specificity: 90%-96%
Positive Predictive Value: 93%-97%
Negative Predictive Value: 87%-94%

Performance for Severity Assessment (cross-validation)
AUC-ROC: 0.97-0.99
Sensitivity: 93%-99%
Specificity: 53%-75%
Positive Predictive Value: 59%-72%
Negative Predictive Value: 94%-99%

If appendix not visualized or USG unavailable (n=144 test cohort):
AUC-ROC: 0.93-0.99
Threshold: >=44% | High Likelihood
Specificity: (93-100)% | PPV: (84-100)%
Sensitivity: (62-86)% | NPV: (80-93)%
Threshold: less than 25%
Specificity: (83-95)% | PPV: (72-92)%
Sensitivity: (84-98)% | NPV: (91-99)%
(Values are reported from cross-validation)

Conceptualized, Designed, and Developed by:
Dr. Anup Thapa Kshetri, MBBS
Founder and Lead: Dharma-AI.org

Sincere gratitude to the team of DharmaAI and my dear friends:

Er. Subash Pahari: Web-app Development and Deployment
Thank you for introducing me to the wonders of codes and AI, and for your invaluable mentorship and support.
You are not just a great friend, but also a guiding force in this project.

Mr. Amrit Neupane: Logo Design | UI/UX Design

×

Prediction Explanation