(BSPOT11) Using Electronic Health Record Data and Neural Network to Predict 90-day Readmission after Peripheral Vascular Intervention: Can it Match and Augment the Capacity of Clinical Registry?

Background: Neural network (NN) models based on claims or registry data often have similar or only slightly better performance than regression models (with area under the curve (AUC) 0.6-0.75), possibly due to limited information available. Electronic health record (EHR) data contain rich information from office settings, lab results, and prescription in addition to diagnoses and procedures in hospital settings, which may help improve predictive performance.

Objectives: To determine whether using NN based on EHR can match or exceed and potentially augment the capacity of registry data predicting 90-day readmission after device-based peripheral vascular intervention (PVI).

Methods: This observational cohort study used linked Vascular Quality Initiative (VQI) registry and INSIGHT Clinical Research Network, aggregated EHR data from multiple academic institutions in New York City. We included 2348 patients undergoing PVI for occlusive disease during 1/1/2013-9/30/2021. The outcome variable was 90-day readmission. In addition to basic patient characteristics (e.g., age, sex, BMI), independent variables from registry data were disease, procedural, and device characteristics, comorbidities, and preoperative and discharge statin and antiplatelet medication. Independent variables from EHR data were diagnoses, procedures, lab results outside normal range, and prescriptions from one year prior to the procedure until discharge. The performance of multiple methods of using EHR variables and NN models were assessed by AUC: 1) raw codes and categorized codes; 2) presence within the year, presence by quarter, and weighted sum by quarterly presence; 3) feed-forward NN/multilayer perceptron (MLP) or recurrent neural network (RNN); 4) models with 1, 2, and 3 hidden layers.

Results: The average age of patients was 70 (±11) years, 44% were male. 832(35%) patients were readmitted within 90 days. The model based on registry data alone has an AUC of 0.721 (SD 0.004). Models based on EHR data were able to match up registry-based model performance (AUC: 0.732-0.742), with MLP using quarterly variables of categorized codes performing best (AUC 0.742, SD 0.003). Increasing the number of hidden layers did not improve prediction performance. Models based on combined EHR and registry data have improved performance (AUC 0.750-0.759). The performance improvement mostly came from adding lab results and prescription data from EHR (AUC 0.752-0.757).

Conclusions: The addition of variables directly derived from EHR data was associated with a small improvement in NN model performance in predicting 90-day readmission after PVI. The set of independent variables that were most helpful from EHR in addition to registry collected data were lab and prescription data.