Improving Clinician Performance in Classification of EEG Patterns on the Ictal-Interictal-Injury Continuum using Interpretable Machine Learning

TL;DR

An interpretable deep-learning system that accurately classifies six patterns of potentially harmful EEG activity - seizure, lateralized periodic discharges (LPDs), generalized periodic discharges (GPDs), lateralized rhythmic delta activity (LRDA), generalized rhythmic delta activity (GRDA), and other patterns - while providing faithful case-based explanations of its predictions is developed.

Abstract

In intensive care units (ICUs), critically ill patients are monitored with electroencephalograms (EEGs) to prevent serious brain injury. The number of patients who can be monitored is constrained by the availability of trained physicians to read EEGs, and EEG interpretation can be subjective and prone to inter-observer variability. Automated deep learning systems for EEG could reduce human bias and accelerate the diagnostic process. However, black box deep learning models are untrustworthy, difficult to troubleshoot, and lack accountability in real-world applications, leading to a lack of trust and adoption by clinicians. To address these challenges, we propose a novel interpretable deep learning model that not only predicts the presence of harmful brainwave patterns but also provides high-quality case-based explanations of its decisions. Our model performs better than the corresponding black box model, despite being constrained to be interpretable. The learned 2D embedded space provides the first global overview of the structure of ictal-interictal-injury continuum brainwave patterns. The ability to understand how our model arrived at its decisions will not only help clinicians to diagnose and treat harmful brain activities more accurately but also increase their trust and adoption of machine learning models in clinical practice; this could be an integral component of the ICU neurologists' standard workflow.

Figures (9)

• Figure 1: Upper: model architecture. Input sample $x$ is passed through a feature extractor $f()$, and a prototype layer $g()$(as in chen2019this). The prototype layer calculates angular distances (as in donnelly2022deformable) between the sample feature and the prototypes. The angular distances are multiplied with class affinity to generate the logits (class scores). The softmax calculation converts the logits into prediction probabilities. Lower: three different ways an end user can see how the model reasons about the test sample: (a) Latent space explanations. (b) Decision space explanations. (c) Scoring system explanations.
• Figure 1: This is a detailed architecture plot of the ProtoPMed-EEG, indicating output dimensions of key components of the model. The check marks ($\checkmark$) in the figure indicate the class connection of each prototype, each single-class prototype has one check mark, and each dual-class prototype has two check marks. The colored heatmap represents class affinity values.
• Figure 2: Snapshot of the graphical user interface (GUI) of the interpretable system. The GUI integrates the three explanation modes detailed in Figure \ref{['fig:3modes_of_interp']}. This snapshot has minor simplifications for ease of reading. Full details and further information on the GUI layout can be found in Appendix \ref{['app:gui']}.
• Figure 2: The graphical user interface (GUI) of the interpretable system. On the left top panel is the 2D embedding map, with each dot representing one EEG sample. Dots can be displayed with shading according to 9 different available schemes (human majority, model prediction, model uncertainty, Seizure burden, etc.). A user can click on the map to select any sample of interest; the 3 nearest prototypes are displayed on the right, ranked according to similarity score (SIM). For each sample/prototype, 10 seconds of the EEG and a 10-minute spectrogram (centered on the 10-second EEG segment) are displayed with human votes and model predictions shown on top of the EEG; a pie chart is provided to visualize the class distribution according to the model or human votes, depending on the selected color scheme. For each prototype, under its pie chart, we also list the values of three terms: similarity score (SIM), class connection (AFF), and class contribution score (SCORE).
• Figure 3: The user study interface when AI assistance is provided. From left to right: buttons to select the EEG pattern category; the EEG sample to be categorized; a comparable EEG prototype provided by the model; a bar chart of class predictions from the model.