Skip to main content Skip to main navigation


Predicting Hemodynamic and Pulmonary Decompensation with Deep Neural Networks: Performance and Explainability

Johannes Rust; Christian Mandel; Kathrin Stich; Serge Autexier
In: 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC-2024), Orlando, USA, 2024.


Predicting the deterioration of patients’ hemodynamic and pulmonary decompensation state while being treated on an intensive care unit is demanding for the medical staff involved. Recent developments in artificial intelligence (AI) show the potential to support physicians in this safety-critical task. This work presents Transformer networks that can predict the hemodynamic and pulmonary decompensation score of a patient up to 24h in advance with mean errors of 3.73% and 2.2% of the maximum achievable score respectively. In an effort towards explainable AI, we analyze the feature attribution of the trained models using Shapley values and present insights to the patterns and relationships learned by the networks. It is shown which medical variables were deemed relevant and when they affect the predictions the most. New techniques to aggregate and plot feature attribution values for time series data are used, extending those from other popular Shapley value packages. Clinical relevance—Physicians working in the field of intensive care and emergency medicine are confronted with hemodynamic and pulmonary decompensation each day. As decompensation progresses, physiological mechanisms are no longer sufficient to prevent severe cardiovascular and respiratory failure, permanent organ damages or death. Prediction and timely detection of these two critical clinical conditions are essential to initiate medical intervention, in order to counteract further progression and to reduce mortality.