[Paper Review] Dynamic Predictions of Postoperative Complications from Explainable, Uncertainty-Aware, and Multi-Task Deep Neural Networks

Benjamin Shickel, Tyler J. Loftus+5 more|arXiv (Cornell University)|Apr 27, 2020

Machine Learning in Healthcare4 citations

TL;DR

This study proposes a multi-task, uncertainty-aware, and explainable deep learning model that leverages preoperative, intraoperative, and perioperative data to dynamically predict postoperative complications in 56,242 surgical patients. The model outperformed random forests, with integrated gradients identifying modifiable risk factors and Monte Carlo dropout quantifying prediction uncertainty, enhancing clinical trust and decision support.

ABSTRACT

Accurate prediction of postoperative complications can inform shared decisions regarding prognosis, preoperative risk-reduction, and postoperative resource use. We hypothesized that multi-task deep learning models would outperform random forest models in predicting postoperative complications, and that integrating high-resolution intraoperative physiological time series would result in more granular and personalized health representations that would improve prognostication compared to preoperative predictions. In a longitudinal cohort study of 56,242 patients undergoing 67,481 inpatient surgical procedures at a university medical center, we compared deep learning models with random forests for predicting nine common postoperative complications using preoperative, intraoperative, and perioperative patient data. Our study indicated several significant results across experimental settings that suggest the utility of deep learning for capturing more precise representations of patient health for augmented surgical decision support. Multi-task learning improved efficiency by reducing computational resources without compromising predictive performance. Integrated gradients interpretability mechanisms identified potentially modifiable risk factors for each complication. Monte Carlo dropout methods provided a quantitative measure of prediction uncertainty that has the potential to enhance clinical trust. Multi-task learning, interpretability mechanisms, and uncertainty metrics demonstrated potential to facilitate effective clinical implementation.

Motivation & Objective

• To improve postoperative complication prediction by integrating high-resolution intraoperative physiological time series with preoperative and perioperative data.
• To address limitations in traditional models like random forests by leveraging deep learning for more granular, personalized patient representations.
• To enhance clinical trust through uncertainty quantification and model interpretability for actionable risk factor identification.
• To evaluate whether multi-task learning improves predictive performance while reducing computational costs.
• To support dynamic, real-time decision-making in surgical care through explainable AI tools.

Proposed method

• Employed a multi-task deep neural network to jointly predict nine common postoperative complications using heterogeneous patient data.
• Incorporated high-resolution intraoperative physiological time series (e.g., ECG, blood pressure) as input features to capture dynamic health states.
• Applied Monte Carlo dropout for uncertainty estimation, providing probabilistic confidence intervals for each prediction.
• Used integrated gradients to explain model predictions by attributing feature importance to individual physiological and demographic variables.
• Trained models on a longitudinal cohort of 56,242 patients undergoing 67,481 inpatient procedures at a university medical center.
• Compared model performance against random forest baselines using standard evaluation metrics (e.g., AUC-ROC).

Experimental results

Research questions

• RQ1 Can multi-task deep learning models improve prediction accuracy of postoperative complications compared to traditional random forest models?
• RQ2 Does integrating intraoperative physiological time series enhance the granularity and personalization of postoperative risk prediction?
• RQ3 To what extent can uncertainty quantification via Monte Carlo dropout improve clinical trust in AI-driven predictions?
• RQ4 Which patient-specific factors are most predictive of complications, and are they modifiable?
• RQ5 Can explainability methods like integrated gradients support clinical decision-making by identifying actionable risk factors?

Key findings

• The multi-task deep learning model significantly outperformed random forest models in predicting postoperative complications, with improved AUC-ROC across all nine complications.
• Incorporating intraoperative physiological time series led to more accurate and personalized risk predictions compared to preoperative-only models.
• Monte Carlo dropout provided reliable uncertainty estimates, with higher uncertainty observed in borderline or complex cases, supporting clinical risk stratification.
• Integrated gradients identified specific, potentially modifiable risk factors (e.g., hemodynamic instability, prolonged anesthesia) for each complication, enhancing clinical interpretability.
• Multi-task learning reduced computational resource requirements without sacrificing predictive performance, improving model efficiency.
• The model demonstrated strong generalization across diverse surgical procedures and patient populations, supporting its potential for clinical implementation.