A supervised, externally validated machine learning model for artifact and drainage detection in high-resolution intracranial pressure monitoring data.

OBJECTIVE: In neurocritical care, data from multiple biosensors are continuously measured, but only sporadically acknowledged by the attending physicians. In contrast, machine learning (ML) tools can analyze large amounts of data continuously, taking advantage of underlying information. However, the performance of such ML-based solutions is limited by different factors, for example, by patient motion, manipulation, or, as in the case of external ventricular drains (EVDs), the drainage of CSF to control intracranial pressure (ICP). The authors aimed to develop an ML-based algorithm that automatically classifies normal signals, artifacts, and drainages in high-resolution ICP monitoring data from EVDs, making the data suitable for real-time artifact removal and for future ML applications. METHODS: In their 2-center retrospective cohort study, the authors used labeled ICP data from 40 patients in the first neurocritical care unit (University Hospital Zurich) for model development. The authors created 94 descriptive features that were used to train the model. They compared histogram-based gradient boosting with extremely randomized trees after building pipelines with principal component analysis, hyperparameter optimization via grid search, and sequential feature selection. Performance was measured with nested 5-fold cross-validation and multiclass area under the receiver operating characteristic curve (AUROC). Data from 20 patients in a second, independent neurocritical care unit (Charité - Universitätsmedizin Berlin) were used for external validation with bootstrapping technique and AUROC. RESULTS: In cross-validation, the best-performing model achieved a mean AUROC of 0.945 (95% CI 0.92-0.969) on the development dataset. On the external validation dataset, the model performed with a mean AUROC of 0.928 (95% CI 0.908-0.946) in 100 bootstrapping validation cycles to classify normal signals, artifacts, and drainages. CONCLUSIONS: Here, the authors developed a well-performing supervised model with external validation that can detect normal signals, artifacts, and drainages in ICP signals from patients in neurocritical care units. For future analyses, this is a powerful tool to discard artifacts or to detect drainage events in ICP monitoring signals.

Machine learning using multimodal and autonomic nervous system parameters predicts clinically apparent stroke-associated pneumonia in a development and testing study.

BACKGROUND: Stroke-associated pneumonia (SAP) is a preventable determinant for poor outcome after stroke. Machine learning (ML) using large-scale clinical data warehouses may be able to predict SAP and identify patients for targeted interventions. The aim of this study was to develop a prediction model for identifying clinically apparent SAP using automated ML. METHODS: The ML model used clinical and laboratory parameters along with heart rate (HR), heart rate variability (HRV), and blood pressure (BP) values obtained during the first 48 h after stroke unit admission. A logistic regression classifier was developed and internally validated with a nested-cross-validation (nCV) approach. For every shuffle, the model was first trained and validated with a fixed threshold for 0.9 sensitivity, then finally tested on the out-of-sample data and benchmarked against a widely validated clinical score (A2DS2). RESULTS: We identified 2390 eligible patients admitted to two-stroke units at Charité between October 2020 and June 2023, of whom 1755 had all parameters available. SAP was diagnosed in 96/1755 (5.5%). Circadian profiles in HR, HRV, and BP metrics during the first 48 h after admission exhibited distinct differences between patients with SAP diagnosis vs. those without. CRP, mRS at admission, leukocyte count, high-frequency power in HRV, stroke severity at admission, sex, and diastolic BP were identified as the most informative ML features. We obtained an AUC of 0.91 (CI 0.88-0.95) for the ML model on the out-of-sample data in comparison to an AUC of 0.84 (CI 0.76-0.91) for the previously established A2DS2 score (p < 0.001). The ML model provided a sensitivity of 0.87 (CI 0.75-0.97) with a corresponding specificity of 0.82 (CI 0.78-0.85) which outperformed the A2DS2 score for multiple cutoffs. CONCLUSIONS: Automated, data warehouse-based prediction of clinically apparent SAP in the stroke unit setting is feasible, benefits from the inclusion of vital signs, and could be useful for identifying high-risk patients or prophylactic pneumonia management in clinical routine.

Datawarehouse-enabled quality control of atrial fibrillation detection in the stroke unit setting.

Objective: (1) To assess the accuracy of a standard operating procedure (SOP) regarding the utilization of atrial fibrillation (AF) alarms in everyday clinical practice, and (2) to evaluate the performance of automated continuous surveillance for atrial fibrillation (AF) in hospitalized acute stroke patients. Design: Retrospective cohort study. Setting: Two stroke units from two tertiary care hospitals in Berlin, Germany. Participants: We identified 635 patients with ischemic stroke diagnosis for the time period between 01. January and 30. September 2021 of which 176 patients had recorded AF alarms during monitoring. Of those, 115 patients were randomly selected for evaluation. After excluding 6 patients with hemorrhagic stroke in their records, 109 patients (mean age: 79.1 years, median NIHSS at admission: 6, 57% female) remained for analysis. Intervention: Using a clinical data warehouse for comprehensive data storage we retrospectively downloaded and visualized ECG data segments of 65 s duration around the automated AF alarms. We restricted the maximum number of ECG segments to ten per patient. Each ECG segment plot was uploaded into a REDCap database and categorized as either AF, non-AF or artifact by manual review. Atrial flutter was subsumed as AF. These classifications were then matched with 1) medical history and known diseases before stroke, 2) discharge diagnosis, and 3) recommended treatment plan in the medical history using electronic health records. Main outcome measures: The primary outcome was the proportion of previously unknown AF diagnoses correctly identified by the monitoring system but missed by the clinical team during hospitalization. Secondary outcomes included the proportion of patients in whom a diagnosis of AF would likely have led to anticoagulant therapy. We also evaluated the accuracy of the automated detection system in terms of its positive predictive value (PPV). Results: We evaluated a total of 717 ECG alarm segments from 109 patients. In 4 patients (3.7, 95% confidence interval [CI] 1.18-9.68%) physicians had missed AF despite at least one true positive alarm. All four patients did not receive long-term secondary prevention in form of anticoagulant therapy. 427 out of 717 alarms were rated true positives, resulting in a positive predictive value of 0.6 (CI 0.56-0.63) in this cohort. Conclusion: By connecting a data warehouse, electronic health records and a REDCap survey tool, we introduce a path to assess the monitoring quality of AF in acute stroke patients. We find that implemented standards of procedure to detect AF during stroke unit care are effective but leave room for improvement. Such data warehouse-based concepts may help to adjust internal processes or identify targets of further investigations.

Data lake-driven analytics identify nocturnal non-dipping of heart rate as predictor of unfavorable stroke outcome at discharge.

BACKGROUND: Post-stroke heart rate (HR) and heart rate variability (HRV) changes have been proposed as outcome predictors after stroke. We used data lake-enabled continuous electrocardiograms to assess post-stroke HR and HRV, and to determine the utility of HR and HRV to improve machine learning-based predictions of stroke outcome. METHODS: In this observational cohort study, we included stroke patients admitted to two stroke units in Berlin, Germany, between October 2020 and December 2021 with final diagnosis of acute ischemic stroke or acute intracranial hemorrhage and collected continuous ECG data through data warehousing. We created circadian profiles of several continuously recorded ECG parameters including HR and HRV parameters. The pre-defined primary outcome was short-term unfavorable functional outcome after stroke indicated through modified Rankin Scale (mRS) score of > 2. RESULTS: We included 625 stroke patients, 287 stroke patients remained after matching for age and National Institute of Health Stroke Scale (NIHSS; mean age 74.5 years, 45.6% female, 88.9% ischemic, median NIHSS 5). Both higher HR and nocturnal non-dipping of HR were associated with unfavorable functional outcome (p < 0.01). The examined HRV parameters were not associated with the outcome of interest. Nocturnal non-dipping of HR ranked highly in feature importance of various machine learning models. CONCLUSIONS: Our data suggest that a lack of circadian HR modulation, specifically nocturnal non-dipping, is associated with short-term unfavorable functional outcome after stroke, and that including HR into machine learning-based prediction models may lead to improved stroke outcome prediction.