A recurrent machine learning model predicts intracranial hypertension in neurointensive care patients

Abstract

The evolution of intracranial pressure (ICP) of critically ill patients admitted to a neurointensive care unit (ICU) is difficult to predict. Besides the underlying disease and compromised intracranial space, ICP is affected by a multitude of factors, many of which are monitored on the ICU, but the complexity of the resulting patterns limits their clinical use. This paves the way for new machine learning (ML) techniques to assist clinical management of patients undergoing invasive ICP monitoring independent of the underlying disease. An institutional cohort (ICP-ICU) of patients with invasive ICP monitoring (n = 1346) was used to train recurrent ML models to predict the occurrence of ICP increases of ≥ 22mmHg over a long (> 2 hours) time period in the upcoming hours. External validation was performed on patients undergoing invasive ICP measurement in two publicly available datasets (Medical Information Mart for Intensive Care (MIMIC, n = 998) and eICU Collaborative Research Database (eICU, n = 1634)). Different distances (1h-24 h) between prediction time point and upcoming critical phase were evaluated, demonstrating a decrease in performance but still robust AUC-ROC with larger distances (24 h AUC-ROC: ICP-ICU 0.826 ± 0.0071, MIMIC 0.836 ± 0.0063, eICU 0.779 ± 0.0046, 1 h AUC-ROC: ICP-ICU 0.982 ± 0.0008, MIMIC 0.965 ± 0.0010, eICU 0.941 ± 0.0025). The model operates on sparse hourly data and is stable in handling variable input lengths and missingness through its nature of recurrence and internal memory. Calculation of gradient-based feature importance revealed individual underlying decisions for our Long Short Time Memory (LSTM) based model and thereby provided improved clinical interpretability. Recurrent ML models have the potential to be an effective tool for the prediction of ICP increases with high translational potential.

Development of Traumatic Brain Injury Associated Intracranial Hypertension Prediction Algorithms: A Narrative Review

“…These efforts can be broadly classified into two related but distinct approaches: 1) ICP forecasting, which involves the development of algorithms designed to predict all future ICP values, and 2) tIH prediction, which involves the development of algorithms designed to specifically provide warning of an impending tIH event. Despite steady advances in both ICP forecasting and tIH prediction algorithm development, these, until very recently, 29 have been limited to the analysis of retrospective data. Several factors have been limiting the translation of these algorithms into clinical practice, including the generalizability-limiting heuristic nature of their development; high feature and/or pre-processing requirements, which results in computational demands that preclude real-time signal processing; and/or the need for a prolonged period of data collection prior to first prediction.…”

“…Most recently, Schweingruber and associates published details of an RNN-LSTM tIH prediction algorithm that can be operationalized for use in prospective clinical studies. 29 The tIH definitions used for training the algorithm were termed short and long critical phases. A short critical phase was defined as any instance of ICP >22 mm Hg during 1 or 2 consecutive-hour blocks, whereas a long critical phase was defined as any instance of ICP >22 mm Hg during 2 or more consecutive-hour blocks.…”

“…A short critical phase was defined as any instance of ICP >22 mm Hg during 1 or 2 consecutive-hour blocks, whereas a long critical phase was defined as any instance of ICP >22 mm Hg during 2 or more consecutive-hour blocks. 29 The time horizon of the algorithm was 1 h. The RNN-LSTM algorithm was trained using an 80%:20% train:test ratio of averaged vital sign wave data and a variety of other numeric data points. The source of the training data was a mixed cohort of ICP-monitored patients (TBI, subarachnoid hemorrhage, stroke, etc.).…”

“…On external validation testing using MIMIC-III data, the AUROC was 0.948 ± 0.0025 and using the eICU data set, 0.903 ± 0.0033. 29 …”

Prediction and risk assessment of sepsis-associated encephalopathy in ICU based on interpretable machine learning

“…In recent years, machine learning has been widely used in medical research, especially in the prediction of diseases in ICU 18 , 19 . In our study, multidimensional feature data were employed to build several machine learning classifiers for predicting the occurrence of SAE within 24 h of ICU admission, and ultimately to select the optimal one.…”