A hybrid ML-PBPK digital twin framework for clinically interpretable readmission risk and drug exposure stratification in diabetes

Vollständiger Abstract

Worum geht es in dieser Arbeit?

The prediction of whether a diabetic patient will return to the hospital within 30 days is a very difficult clinical situation to navigate. Predicting the chance of a 30- day readmission depends on a wide range of patient-specific risk factors and the variability in how individual patients respond to medication prescribed for them. Traditional machine-learning models are capable of finding patterns in electronic medical records but are typically unable to provide a clear picture of the physiological interactions between the patient and their prescribed drug(s). One way to predict the likelihood of diabetes-related rehospitalization is using a combination of machine-learning risk factors and physiologically based pharmacokinetics (PBPK). An optimized version of the XG Boost machine-learning algorithm was trained on a large in-patient dataset (approximately 100,000 patient encounters) and achieved a predictive model with an area under the receiver operating characteristic curve (AUC) of 0.68 with a recall of 0.60 at clinically meaningful cut-offs. These findings are similar to what has been previously reported in the literature (0.60–0.70) for diabetes-related rehospitalization risk and support the utility of the XG Boost algorithm as a reliable clinical screening tool. Concurrently, multiple PBPK simulations were run to assess the effects chronic renal or hepatic impairments have on PBPK; results show renal impairment results in the highest systemic drug exposures. Predictive risk and simulated exposure will be integrated to yield a patient-specific risk/exposure phenotype and enable clinically meaningful stratification into actionable subgroups. Through this new method, it becomes possible to discern pharmacological vs. non-pharmacological contributors to readmission risk, thus facilitating targeted intervention strategies such as dose modulation, enhanced monitoring, and coordinated care. This work's value lies not in improving prediction accuracy by just a few points, but rather in a new approach to converting static risk prediction into a clinically actionable decision support system based on physiology. The proposed Digital Twin framework offers a scalable path to personalized post-discharge patient management and provides the basis for furthering the implementation of precision medicine in the field of digital health.