AI-Assisted Prediction of Postoperative Physiological Complications Using Multimodal Clinical and Biosignal Data: A Translational Healthcare Framework

This is the authors' abstract. We don't add key points for this paper.

Background: The aim of this study was to develop and internally validate a multimodal clinical and physiological risk prediction model. Methods: Clinical data comprising 11 variables and 12 features extracted from physiological time-series data (48 data points) for heart rate, oxygen saturation, mean arterial pressure and respiratory rate were generated for 4,000 simulated surgical cases. Endpoints relating to respiratory, cardiovascular and renal complications were defined using prespecified stochastic processes involving risk factors and physiological states. The data were divided into training (n=2,400), validation (n=800) and test (n=800) sets. Three models (logistic regression, random forest and histogram gradient boosting) were compared using clinical features alone, physiological features alone and combined multimodal features. Results: The overall event rate in the dataset was 33.7% (1,348 of 4,000 cases). The multimodal logistic regression model selected using the validation set demonstrated a test-set AUROC of 0.7682 (95% bootstrap CI 0.7301–0.8007), an AUPRC of 0.6444 and a Brier score of 0.1763. The AUROCs for logistic regression using clinical features alone and physiological features alone were 0.6965 and 0.7144, respectively. Conclusions: The internal discriminatory ability was increased by combining clinical and physiological variables. However it does not represent performance on real post-operative patients and is not shown to be clinically useful.