Demystifying the Black Box: The Importance of Interpretability of Predictive Models in Neurocritical Care

Moss, Laura; Corsar, David; Shaw, Martin; Piper, Ian; Hawthorne, Christopher

doi:10.1007/s12028-022-01504-4

Cited by 16 publications

(10 citation statements)

References 44 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Accurate risk prediction models have the potential to improve healthcare by directing timely interventions to patients who are most likely to benefit. However, prediction models that cannot scale adequately to large databases or cannot be interpreted and explained will struggle to gain acceptance in clinical practice [Moss et al, 2022]. The current study advances the oblique RSF, an accurate risk prediction model, towards being accurate, scalable, and interpretable.…”

Section: Implications Of Our Resultsmentioning

confidence: 98%

Accelerated and interpretable oblique random survival forests

Jaeger¹,

Welden²,

Kristin³

et al. 2022

Preprint

View full text Add to dashboard Cite

The oblique random survival forest (RSF) is an ensemble supervised learning method for rightcensored outcomes. Trees in the oblique RSF are grown using linear combinations of predictors to create branches, whereas in the standard RSF, a single predictor is used. Oblique RSF ensembles often have higher prediction accuracy than standard RSF ensembles. However, assessing all possible linear combinations of predictors induces significant computational overhead that limits applications to large-scale data sets. In addition, few methods have been developed for interpretation of oblique RSF ensembles, and they remain more difficult to interpret compared to their axis-based counterparts. In this article, we introduce and evaluate a method to increase computational efficiency of the oblique RSF and a method to estimate importance of individual predictor variables with the oblique RSF. Our strategy to reduce computational overhead makes use of Newton-Raphson scoring, a classical optimization technique that we apply to the Cox partial likelihood function within each non-leaf node of decision trees. We estimate the importance of individual predictors for the oblique RSF by negating each coefficient used for the given predictor in linear combinations, and then computing the reduction in out-of-bag accuracy. In general benchmarking experiments, we find that our implementation of the oblique RSF is approximately 450 times faster with equivalent discrimination and superior Brier score compared to existing software for oblique RSFs. We find in simulation studies that 'negation importance' discriminates between relevant and irrelevant predictors more reliably than permutation

show abstract

Section: Implications Of Our Resultsmentioning

confidence: 98%

Accelerated and interpretable oblique random survival forests

Jaeger¹,

Welden²,

Kristin³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…This approach has been previously used to understand machine learning models employed in the context of neurocritical care. 34 The ten variables with the highest mean absolute SHAP values were considered the most influential features in this analysis. The SHAP Python library (version 0.44.0) facilitated the analysis of the most influential features by generating (1) a summary plot of mean absolute SHAP values, (2) scatter plots of SHAP values for each variable, (3) a heatmap of SHAP interaction values, and (4) scatter plots of important interactions.…”

Section: Methodsmentioning

confidence: 99%

Machine Learning Identifies Variation in Timing of Palliative Care Consultations Among Traumatic Brain Injury Patients

Aude,

Vattipally,

Das

et al. 2024

Preprint

View full text Add to dashboard Cite

Background and Objective Timely palliative care involvement offers demonstrable benefits for traumatic brain injury (TBI) patients; however, palliative care consultations (PCCs) are used inconsistently during TBI management. This study aimed to employ advanced machine learning techniques to elucidate the primary drivers of PCC timing variability for TBI patients. Methods Data on admission, hospital course, and outcomes were collected for a cohort of 232 TBI patients who received both PCCs and neurosurgical consultations during the same hospitalization. Principal Component Analysis (PCA) and K-means clustering were used to identify patient phenotypes, which were then compared using Kaplan-Meier analysis. An extreme gradient boosting model (XGBoost) was employed to determine drivers of PCC timing, with model interpretation performed using SHapley Additive exPlanations (SHAP). Results Cluster A (n = 86) consisted mainly of older (median [IQR] = 87 [78, 94] years), White females with mild TBIs and demonstrated the shortest time-to-PCC (2.5 [1.0, 7.0] days). Cluster B (n = 108) also sustained mild TBIs but comprised moderately younger (81 [75, 86] years) married White males with later PCC (5.0 [3.0, 10.8] days). Cluster C (n = 38) represented much younger (46.5 [29.5, 59.8] years), more severely injured, non-White patients with the latest PCC initiation (9.0 [4.2, 17.0] days). The clusters did not differ by discharge disposition (p = 0.4) or frequency inpatient mortality (p > 0.9); however, Kaplan-Meier analysis revealed a significant difference in the time from admission to PCC (p < 0.001), despite no differences in time from admission to mortality (p = 0.18). SHAP analysis of the XGBoost model identified age, sex, and race as the most influential drivers of PCC timing. Conclusions This study highlights crucial disparities in PCC timing for TBI patients and underscores the need for targeted strategies to ensure timely and equitable palliative care integration for this vulnerable population.

show abstract

“…While current ML-based classification systems yield good prediction accuracy, a significant hurdle to their broad application is the lack of attention given by researchers to the problem of model interpretability ( 61 , 68 ). In addition, considerable work is required to address the question of how effectively models can be perceived by humans.…”

Section: Model Trustworthy and Interpretabilitymentioning

confidence: 99%

Comparative effectiveness of explainable machine learning approaches for extrauterine growth restriction classification in preterm infants using longitudinal data

Cho,

Kim,

Kim

et al. 2023

Front. Med.

View full text Add to dashboard Cite

IntroductionPreterm birth is a leading cause of infant mortality and morbidity. Despite the improvement in the overall mortality in premature infants, the intact survival of these infants remains a significant challenge. Screening the physical growth of infants is fundamental to potentially reducing the escalation of this disorder. Recently, machine learning models have been used to predict the growth restrictions of infants; however, they frequently rely on conventional risk factors and cross-sectional data and do not leverage the longitudinal database associated with medical data from laboratory tests.MethodsThis study aimed to present an automated interpretable ML-based approach for the prediction and classification of short-term growth outcomes in preterm infants. We prepared four datasets based on weight and length including weight baseline, length baseline, weight follow-up, and length follow-up. The CHA Bundang Medical Center Neonatal Intensive Care Unit dataset was classified using two well-known supervised machine learning algorithms, namely support vector machine (SVM) and logistic regression (LR). A five-fold cross-validation, and several performance measures, including accuracy, precision, recall and F1-score were used to compare classifier performances. We further illustrated the models’ trustworthiness using calibration and cumulative curves. The visualized global interpretations using Shapley additive explanation (SHAP) is provided for analyzing variables’ contribution to final prediction.ResultsBased on the experimental results with area under the curve, the discrimination ability of the SVM algorithm was found to better than that of the LR model on three of the four datasets with 81%, 76% and 72% in weight follow-up, length baseline and length follow-up dataset respectively. The LR classifier achieved a better ROC score only on the weight baseline dataset with 83%. The global interpretability results revealed that pregnancy-induced hypertension, gestational age, twin birth, birth weight, antenatal corticosteroid use, premature rupture of membranes, sex, and birth length were consistently ranked as important variables in both the baseline and follow-up datasets.DiscussionThe application of machine learning models to the early detection and automated classification of short-term growth outcomes in preterm infants achieved high accuracy and may provide an efficient framework for clinical decision systems enabling more effective monitoring and facilitating timely intervention.

show abstract

Demystifying the Black Box: The Importance of Interpretability of Predictive Models in Neurocritical Care

Cited by 16 publications

References 44 publications

Accelerated and interpretable oblique random survival forests

Accelerated and interpretable oblique random survival forests

Machine Learning Identifies Variation in Timing of Palliative Care Consultations Among Traumatic Brain Injury Patients

Comparative effectiveness of explainable machine learning approaches for extrauterine growth restriction classification in preterm infants using longitudinal data

Contact Info

Product

Resources

About