An Investigation of Applications of Artificial Neural Networks in Medical Prognostics
Ghavami, Peter K.
MetadataShow full item record
During the course of care, patients frequently develop escalating health problems that lead to medical complications, costly treatments, severe pains, disabilities and even death. Predicting such escalations provides the opportunity to apply preventive measures that result in better patient safety, quality of care and lower medical costs; in short, timely prediction can save lives and avoid further medical complications. Prognostics methods using Artificial Neural Networks (ANN) promise to deliver new insights into future patient health status that provide more effective medical treatment during the patient hospital stay. With the advent of smaller, inexpensive sensors and volume of data collected from patients, physicians are challenged with making increasingly analytical decisions from a large set of data that are being collected per patient. This trend is only increasing giving rise to what's known in the industry as the "big data problem": The rate of data accumulation is rising faster than physicians' cognitive capacity to analyze increasingly large data sets to make decisions. The big data problem offers an opportunity for predictive analytics and prognostics. Investigation and development of a methodical framework for medical data prognostics in general and use of committee of algorithms in particular have not been adequately explored. A framework for prediction of patient health status from clinical data is needed to assist physicians in their clinical decision process. This research investigates and contributes to three essential ideas for improving healthcare prognostics through big data analytics: 1) A control system approach to prognostics for prediction, 2) A generalized committee of models framework as prognostics engine, and 3) Study the viability of such framework on a particular clinical case. This research offers three key contributions: First, it develops a control system treatment of medical prognostics and predictive models. The control system development of prognostics combines feed-forward and feedback control mechanisms to create a framework for medical prognostics. This framework introduces a rules-based prognostics engine that uses ANN algorithms to identify patients who develop a particular disease or medical complication. Second, it provides a generalized committee of models framework to predict the patient's medical condition and predict any medical complication from large data sets. The model also provides the strength (or the impact level) of all contributing clinical data to that prediction. The methodology proposes using a multi-algorithm prognostics framework to enhance the accuracy of prediction using four ANN models. The framework introduces a supervisory program, called an oracle to select the most appropriate ensemble of models that best meet the practitioner's desired prediction accuracy. Third, it demonstrates the viability and feasibility of using ANN methods as predictive models in this framework. As part of the demonstration, the research explores building, training and validating four ANN models to predict medical complications from data acquired during 1,073 patients' hospital stay to predict Deep Vein Thrombosis/Pulmonary Embolism (DVT/PE). DVT/PE, is a condition caused by blockage of patient lung vessels by blood clots that initially form in patient's legs. DVT/PE leads to severe pain, loss of lung function and even death. The aim of all three ideas is to improve the physician's ability to make predictive decisions from a vast array of data in order to be proactive and apply preventative medical interventions before complications occur.