Abstract
The Nikshy platform in India helps in monitoring Tuberculosis (TB) cases across over 50,000 health facilities in the country. It is a retrospective reporting tool that records facility incidents after they occur. This has resulted in issues like problems at individual facilities going unnoticed, a difference in reporting between rural facilities and those in urban private hospitals and a total number of TB cases going unreported each month. In order to address these issues, we have proposed a dual-stream graph neural network that provides a facility prediction component along with a diagnosis component for efficient resource allocation and high-risk patient prioritization. The facility prediction component utilizes synthetic data from Nikshy for 50 facilities over a period of 36 months, including TB notifications, treatment initiations and rural facilities. A spatio-temporal graph autoencoder model that combines gated recurrent units (GRU) and graph convolutional networks (GCN) results in a reconstruction mean squared error of 0.032 along with anomaly detection AUC of 0.89 compared to ARIMA’s AUC of 0.71. Anomalies are predicted based on errors exceeding the 95th percentile for 1-2 months’ advance warning on facility issues, where TB notification counts are found to be the most important feature by SHAP analysis. Complementarily, the diagnostic flow processes chest CT scans via lung segmentation and graph-based feature extraction to rapidly identify TB risk via heatmaps for prioritized patients. The system improves public health responsiveness by unifying surveillance silos with interpretable AI for TB control.
Keywords: Graph Neural Networks, Health Equity, Nikshay, Shapley Additive Explanations, Spatio-Temporal Graph Autoencoder, Tuberculosis Surveillance.