Predicting vital sign deviations during surgery from patient monitoring data: developing and validating single-stream deep learning models

Editor—In a joint project between two Swiss universities, we developed machine learning models to predict deviations from upper and lower thresholds across 13 distinct vital signs, as outlined in Table 1. This project had two primary goals: to demonstrate the feasibility of short-term predictions (20 s–20 min) across a wider range of vital signs than previously researched; and to develop models for a future Artificial Intelligence Clinical Decision Support System (AI-CDSS) to be used in patient monitors, relying only on the continuous vital sign data available in these monitors. Such a system could proactively alert clinicians to impending vital sign changes, enabling adjustments based on forecasts rather than alarms. 1 Rajkomar A. Dean J. Kohane I. Machine learning in medicine. N Engl J Med. 2019; 380: 1347-1358 Google Scholar We envision a system that, similar to the common trend displays for the recent past of each vital sign, would provide predictive trend indications. Table 1Performance of the vital sign prediction models (classification task). Numbers are balanced accuracy (=[Sensitivity+Specificity]/2). Predictions with a forecast horizon of seconds use the high-resolution dataset (1 s−1), and predictions in minutes use the 1 min−1 dataset. Values in bold: balanced accuracy 0.50–0.70; values in italic: balanced accuracy 0.71–0.84; values in regular font: balanced accuracy 0.85–1. The vital sign thresholds, for which the models achieved the balanced accuracies shown here, are listed in Supplementary Table S2. BIS, Bispectral Index®; CVP, central venous pressure; etCO2, end-tidal CO2 concentration; HR, heart rate; IABP, invasive arterial blood pressure; N/A, not available; NIABP, noninvasive arterial blood pressure; PR, pulse rate; RR, respiratory rate; SpO2, oxygen saturation; Temp, temperature; TOF%, train-of-four ratio; TV, tidal volume.