Objective

This study was undertaken to develop a machine learning (ML) model to forecast initial seizure onset in neonatal hypoxic-ischemic encephalopathy (HIE) utilizing clinical and quantitative electroencephalogram (QEEG) features.

Methods

We developed a gradient boosting ML model (Neo-GB) that utilizes clinical features and QEEG to forecast time-dependent seizure risk. Clinical variables included cord blood gas values, Apgar scores, gestational age at birth, postmenstrual age (PMA), postnatal age, and birth weight. QEEG features included statistical moments, spectral power, and recurrence quantification analysis (RQA) features. We trained and evaluated Neo-GB on a University of California, San Francisco (UCSF) neonatal HIE dataset, augmenting training with publicly available neonatal electroencephalogram (EEG) datasets from Cork University and Helsinki University Hospitals. We assessed the performance of Neo-GB at providing dynamic and static forecasts with diagnostic performance metrics and incident/dynamic area under the receiver operating characteristic curve (iAUC) analyses. Model explanations were performed to assess contributions of QEEG features and channels to model predictions.

Results

The UCSF dataset included 60 neonates with HIE (30 with seizures). In subject-level static forecasting at 30 min after EEG initiation, baseline Neo-GB without time-dependent features had an area under the receiver operating characteristic curve (AUROC) of .76 and Neo-GB with time-dependent features had an AUROC of .89. In time-dependent evaluation of the initial seizure onset within a 24-h seizure occurrence period, dynamic forecast with Neo-GB demonstrated median iAUC = .79 (interquartile range [IQR] .75-.82) and concordance index (C-index) = .82, whereas baseline static forecast at 30 min demonstrated median iAUC = .75 (IQR .72-.76) and C-index = .69. Model explanation analysis revealed that spectral power, PMA, RQA, and cord blood gas values made the strongest contributions in driving Neo-GB predictions. Within the most influential EEG channels, as the preictal period advanced toward eventual seizure, there was an upward trend in broadband spectral power.

Significance

This study demonstrates an ML model that combines QEEG with clinical features to forecast time-dependent risk of initial seizure onset in neonatal HIE. Spectral power evolution is an early EEG marker of seizure risk in neonatal HIE.

Artificial intelligence and machine learning (AI/ML) is becoming increasingly more accessible to biomedical researchers with significant potential to transform biomedicine through optimization of highly-accurate predictive models and enabling better understanding of disease biology. Automated machine learning (AutoML) in particular is positioned to democratize artificial intelligence (AI) by reducing the amount of human input and ML expertise needed. However, successful translation of AI/ML in biomedicine requires moving beyond optimizing only for prediction accuracy and towards establishing reproducible clinical and biological inferences. This is especially challenging for clinical studies on rare disorders where the smaller patient cohorts and corresponding sample size is an obstacle for reproducible modeling results. Here, we present a model-agnostic framework to reinforce AutoML using strategies and tools of explainable and reproducible AI, including novel metrics to assess model reproducibility. The framework enables clinicians to interpret AutoML-generated models for clinical and biological verifiability and consequently integrate domain expertise during model development. We applied the framework towards spinal cord injury prognostication to optimize the intraoperative hemodynamic range during injury-related surgery and additionally identified a strong detrimental relationship between intraoperative hypertension and patient outcome. Furthermore, our analysis captured how evolving clinical practices such as faster time-to-surgery and blood pressure management affect clinical model development. Altogether, we illustrate how expert-augmented AutoML improves inferential reproducibility for biomedical discovery and can ultimately build trust in AI processes towards effective clinical integration.

[This corrects the article DOI: 10.1371/journal.pone.0265254.].

Biomedical research and clinical practice are in the midst of a transition toward significantly increased use of artificial intelligence (AI) and machine learning (ML) methods. These advances promise to enable qualitatively deeper insight into complex challenges formerly beyond the reach of analytic methods and human intuition while placing increased demands on ethical and explainable artificial intelligence (XAI), given the opaque nature of many deep learning methods. The U.S. National Institutes of Health (NIH) has initiated a significant research and development program, Bridge2AI, aimed at producing new "flagship" datasets designed to support AI/ML analysis of complex biomedical challenges, elucidate best practices, develop tools and standards in AI/ML data science, and disseminate these datasets, tools, and methods broadly to the biomedical community. An essential set of concepts to be developed and disseminated in this program along with the data and tools produced are criteria for AI-readiness of data, including critical considerations for XAI and ethical, legal, and social implications (ELSI) of AI technologies. NIH Bridge to Artificial Intelligence (Bridge2AI) Standards Working Group members prepared the present article to present methods for assessing the AI-readiness of biomedical data and the data standards perspectives and criteria we have developed throughout this program. While the field is rapidly evolving, these criteria are foundational for scientific rigor and the ethical design and application of biomedical AI methods.