Assessment and improvement of predictive value of complexity based features for sepsis prediction

Abstract

Data analytics is not embedded into critical environments such as Intensive Care Unit. It is now realized that there is an untapped potential of Big-data in improving patient care. However, this is not taken into practice due to lack of communication between the doctors and the data-analysts. To bridge this gap, we worked on Intensive Care Unit (ICU) data from two sources, one being numeric data from version 3.1 of the Multiparameter intelligent monitoring in intensive care (MIMIC II) waveform Matched subset and second one being Vital Monitoring data from the Pediatric Intensive Care Unit at the All India Institute of Medical Sciences (AIIMS), New Delhi. This thesis presents unsupervised modelling on Numeric data of MIMIC Matched Subset, called Permutation Distribution Clustering (PDC). It is a complexity-based approach to cluster time series data. Clusters obtained by clustering are validated for their clinical potential and examined in order to determine the enrichment of sepsis. It is calculated using SIRS score which uses numeric and clinical data of MIMIC. This work also involves development of a doctor-friendly dashboard that carries out analytics on the Vital Monitoring data from the Pediatric Intensive Care Unit at the All India Institute of Medical Sciences (AIIMS), New Delhi. It displays key patient indices calculated from the continuous data. The dashboard is updated regularly in an hourly fashion. The dashboard has deployed Quality Check, Basic Analytics and Advanced Analytics modalities. This dashboard is designed for Quality Improvement through display of basic Quality Check mechanisms such as patient information checks, technical anomalies and fidelity of sensor data. It is also designed to give care-related feedback to Intensivists and Nurses.