Developing hospital readmission risk prediction models and remote follow-up system for multiple diseases

The objectives of the project are as follows:

1. Developing readmission risk prediction models for multiple diseases: We will develop statistical machine learning models to predict the risk of readmission initially for the following 8 disease groups: myocardial infarction, chronic obstructive pulmonary disease, heart failure, pneumonia, diabetes mellitus, stroke, dementia, and Alzheimer’s disease. We initially target these diseases to maximize the impact of the project, since these diseases are considered the most contributing ones to the yearly overall unplanned readmission costs in Turkey.

2. Identifying unplanned readmission cases in a personalized and adaptive way: The current practice in the field employs one-fits-all kind of thresholds that purely focus on the number of days that has passed since discharge (e.g., considering all re-hospitalizations within 30 days as “readmission”). Any hospitalization before this threshold is considered as “readmission”, while later hospitalizations are not considered as “readmission”. Such a coarse granularity approach ignores patient status at the time of hospitalization. Besides, the employed thresholds do not rely on patient data and are not well-agreed on among experts. In this project, we will develop a separate set of machine learning models that will mark each hospitalization as “readmission” or not based on patient data.

3. Locating and eliminating data entry errors: Even though hospital records are stored in a computerized manner, in many countries, the reliability of some fields in patient records is questionable. During our preliminary study of the underlying datasets, as well as based on our interviews with the healthcare professionals, we observe that during data entry, due to high work overload and time pressure, some healthcare professionals may just choose options that are most convenient and least time consuming on the software interface that they are interacting with. In healthcare systems (e.g., in Turkey) with none-to-minimal auditing practices, such data entry errors are commonplace. Data entry errors should be eliminated before training any statistical models on the data. Manually identifying and eliminating such errors is time-consuming and resource-intensive. As part of this project, we will develop supervised and unsupervised machine learning models to locate and eliminate data entry errors caused by inattentive health care professionals.      

4. Developing a patient discharge decision support system based on readmission risk: We will develop a clinical decision support system and integrate the above readmission risk prediction models as a core component of it to aid clinicians so that they may make a more informed discharge decision based on patient data.

5. Developing a remote follow-up program integrated with readmission risk prediction models and connected to the workflow of healthcare providers: We will develop a remote patient monitoring program that will allow care providers to follow-up with their patients after they are discharged. In particular, participating patients will temporarily be provided with a set of portable measurement devices, e.g., blood sugar measurement tool, blood pressure monitor, etc. at the time of discharge. Then, participating patients will be asked to periodically communicate their measurements through a mobile app or web-based tool to their health care provider. The previously built readmission risk models will continuously evaluate the stream of remote measurements provided by each patient. If the readmission risk is found to be high, then both the healthcare provider and the patient will be notified. Accordingly, a physician will get in touch with the patient and schedule a visit or a planned admission based on the readmission risk score, the availability of hospital bed space, as well as the availability of the patient.