Sabari Giri S Murugan¹ , R Sudhakaran²

Abstract

Purpose: A scalable patient similarity network was developed based on the Medical Information Mart for Intensive Care. Methods: The unit consists of 40,000 patients as network nodes and 100 million weighted edges, where a connection between a patient and a doctor represents diagnoses, laboratory measurements, vital signs, and medication history. The propagation of similarity was performed using the Random Walk with Restart algorithm with a restart probability of 0.35, utilizing a hybrid system of central processing unit clusters and graphics processing unit acceleration, which achieved a speedup of 8.2 times. The network also exhibited small-world properties, with a diameter of 7 and a modularity of 0.42. Results: Predictive analysis showed that the Random Walk with Restart algorithm and a hybrid Random Walk with Restart and Graph Neural Network ensemble had areas under the receiver operating characteristic curves of 0.91 and 0.94, respectively, which are better than those of the conventional models. Conclusion: This has been made possible by incremental updates that enable near-real-time incorporation of new patients, providing an efficient, scalable, and clinically interpretable framework for sepsis risk prediction.