Enriching the feature space for early identification of patient deterioration
Principal Investigator: Mr Ritankar Das
Approved Research ID: 52414
Approval date: June 17th 2020
Electronically available medical data, including patient demographics and online reporting and surveys provide a valuable, but underutilized, source of patient information. Dascena's technology seeks to collapse the vast amounts of patient data into concise care recommendations, such as risk of patient deterioration, to aid with clinical decision making, and to improve patient care and outcomes. Our machine-learning based technology examines correlations between patient data, such as demographics and vital signs to extract additional information about patient trends and status. These inputs currently include basic information routinely entered into a patient's electronic health record (EHR) such as age and vital signs. We believe that the potential accuracy and detection sensitivity can be improved by enriching the data inputs. The data available at UK BioBank represents a valuable opportunity for us to determine which types of new patient data inputs can be used to develop, train and test a machine learning-based technology that provides accurate predictions of patient deterioration. The goal of this 12-month project is to use the patient data obtained from the UK BioBank database to learn relationships between different pieces of demographic information and patient characteristics. Then, using these relationships, we can infer these pieces of information when some of them are missing. This will enable us to increase the value of clinical datasets for research and the development of new tools for predicting developments in patient health. We do this by learning how various patient characteristics and measurements typically precede events of clinical relevance. In this project, we aim to improve the quality of 48-hour mortality prediction. These predictions can form alerts to clinicians, so that they have advance warning of patient deterioration, and can intervene to improve patient outcomes. While the scope of this proposal is limited to this particular application, we believe that the work it entails will teach us more generally about the possibilities of 'filling-in' information that is unavailable in one dataset, but may be available in another. As the predictions we make about patient health depend on the data which are available, this work could benefit many adjacent areas of research, and improve applications beyond 48-hour mortality prediction.