ABSTRACT
OBJECTIVE: The role of surgery in spontaneous intracerebral hemorrhage (SICH) remains controversial. We aimed to use explainable machine learning (ML) combined with propensity-score matching to investigate the effects of surgery and identify subgroups of patients with SICH who may benefit from surgery in an interpretable fashion. METHODS: We conducted a retrospective study of a cohort of 282 patients aged ≥21 years with SICH. ML models were developed to separately predict for surgery and surgical evacuation. SHapley Additive exPlanations (SHAP) values were calculated to interpret the predictions made by ML models. Propensity-score matching was performed to estimate the effect of surgery and surgical evacuation on 90-day poor functional outcomes (PFO). RESULTS: Ninety-two patients (32.6%) underwent surgery, and 57 patients (20.2%) underwent surgical evacuation. A total of 177 patients (62.8%) had 90-day PFO. The support vector machine achieved a c-statistic of 0.915 when predicting 90-day PFO for patients who underwent surgery and a c-statistic of 0.981 for patients who underwent surgical evacuation. The SHAP scores for the top 5 features were Glasgow Coma Scale score (0.367), age (0.214), volume of hematoma (0.258), location of hematoma (0.195), and ventricular extension (0.164). Surgery, but not surgical evacuation of the hematoma, was significantly associated with improved mortality at 90-day follow-up (odds ratio, 0.26; 95% confidence interval, 0.10-0.67; P = 0.006). CONCLUSIONS: Explainable ML approaches could elucidate how ML models predict outcomes in SICH and identify subgroups of patients who respond to surgery. Future research in SICH should focus on an explainable ML-based approach that can identify subgroups of patients who may benefit functionally from surgical intervention.
Subject(s)
Cerebral Hemorrhage , Support Vector Machine , Humans , Retrospective Studies , Propensity Score , Cerebral Hemorrhage/complications , Hematoma/surgery , Treatment OutcomeABSTRACT
OBJECTIVE: Intracranial pressure (ICP) monitoring is a widely utilized and essential tool for tracking neurosurgical patients, but there are limitations to the use of a solely ICP-based paradigm for guiding management. It has been suggested that ICP variability (ICPV), in addition to mean ICP, may be a useful predictor of neurological outcomes, as it represents an indirect measure of intact cerebral pressure autoregulation. However, the current literature regarding the applicability of ICPV shows conflicting associations between ICPV and mortality. Thus, the authors aimed to investigate the effect of ICPV on intracranial hypertensive episodes and mortality using the eICU Collaborative Research Database version 2.0. METHODS: The authors extracted from the eICU database 1,815,676 ICP readings from 868 patients with neurosurgical conditions. ICPV was computed using two methods: the rolling standard deviation (RSD) and the absolute deviation from the rolling mean (DRM). An episode of intracranial hypertension was defined as at least 25 minutes of ICP > 22 mm Hg in any 30-minute window. The effects of mean ICPV on intracranial hypertension and mortality were computed using multivariate logistic regression. A recurrent neural network with long short-term memory was used for time-series predictions of ICP and ICPV to prognosticate future episodes of intracranial hypertension. RESULTS: A higher mean ICPV was significantly associated with intracranial hypertension using both ICPV definitions (RSD: aOR 2.82, 95% CI 2.07-3.90, p < 0.001; DRM: aOR 3.93, 95% CI 2.77-5.69, p < 0.001). ICPV was significantly associated with mortality in patients with intracranial hypertension (RSD: aOR 1.28, 95% CI 1.04-1.61, p = 0.026, DRM: aOR 1.39, 95% CI 1.10-1.79, p = 0.007). In the machine learning models, both definitions of ICPV achieved similarly good results, with the best F1 score of 0.685 ± 0.026 and an area under the curve of 0.980 ± 0.003 achieved with the DRM definition over 20 minutes. CONCLUSIONS: ICPV may be useful as an adjunct for the prognostication of intracranial hypertensive episodes and mortality in neurosurgical critical care as part of neuromonitoring. Further research on predicting future intracranial hypertensive episodes with ICPV may help clinicians react expediently to ICP changes in patients.
Subject(s)
Brain Injuries, Traumatic , Intracranial Hypertension , Humans , Intracranial Pressure/physiology , Critical Illness , Monitoring, Physiologic , Logistic Models , Intracranial Hypertension/diagnosis , Intracranial Hypertension/etiology , Brain Injuries, Traumatic/surgeryABSTRACT
OBJECTIVE: This study aims to develop and compare the use of deep neural networks (DNN) and support vector machines (SVM) to clinical prognostic scores for prognosticating 30-day mortality and 90-day poor functional outcome (PFO) in spontaneous intracerebral haemorrhage (SICH). MATERIALS AND METHODS: We conducted a retrospective cohort study of 297 SICH patients between December 2014 and May 2016. Clinical data was collected from electronic medical records using standardized data collection forms. The machine learning workflow included imputation of missing data, dimensionality reduction, imbalanced-class correction, and evaluation using cross-validation and comparison of accuracy against clinical prognostic scores. RESULTS: 32 (11%) patients had 30-day mortality while 177 (63%) patients had 90-day PFO. For prognosticating 30-day mortality, the class-balanced accuracies for DNN (0.875; 95% CI 0.800-0.950; McNemar's p-value 1.000) and SVM (0.848; 95% CI 0.767-0.930; McNemar's p-value 0.791) were comparable to that of the original ICH score (0.833; 95% CI 0.748-0.918). The c-statistics for DNN (0.895; DeLong's p-value 0.715), and SVM (0.900; DeLong's p-value 0.619), though greater than that of the original ICH score (0.862), were not significantly different. For prognosticating 90-day PFO, the class-balanced accuracies for DNN (0.853; 95% CI 0.772-0.934; McNemar's p-value 0.003) and SVM (0.860; 95% CI 0.781-0.939; McNemar's p-value 0.004) were better than that of the ICH-Grading Scale (0.706; 95% CI 0.600-0.812). The c-statistic for SVM (0.883; DeLong's p-value 0.022) was significantly greater than that of the ICH-Grading Scale (0.778), while the c-statistic for DNN was 0.864 (DeLong's p-value 0.055). CONCLUSION: We showed that the SVM model performs significantly better than clinical prognostic scores in predicting 90-day PFO in SICH.