This EEG Looks Like These EEGs: Interpretable Interictal Epileptiform Discharge Detection With ProtoEEG-kNN

TL;DR

The paper tackles automated detection of interictal epileptiform discharges (IEDs) in EEG and the need for interpretable AI in clinical practice. It introduces ProtoEEG-kNN, an interpretable, prototype-based framework that learns a meaningful embedding via a ProtoPNet-inspired space augmented with an interpretable similarity metric over interpretable statistics (ISFs) and channel-wise weights, then uses a top-$k$ nearest-neighbors rule with $k=10$ over training samples. Results show state-of-the-art accuracy on a large, multi-center EEG dataset, supported by a user study indicating that explanations align with expert reasoning. The work provides a practical path for integrating human-understandable ML into epilepsy diagnostics and highlights the need for external validation across diverse patient populations.

Abstract

The presence of interictal epileptiform discharges (IEDs) in electroencephalogram (EEG) recordings is a critical biomarker of epilepsy. Even trained neurologists find detecting IEDs difficult, leading many practitioners to turn to machine learning for help. While existing machine learning algorithms can achieve strong accuracy on this task, most models are uninterpretable and cannot justify their conclusions. Absent the ability to understand model reasoning, doctors cannot leverage their expertise to identify incorrect model predictions and intervene accordingly. To improve the human-model interaction, we introduce ProtoEEG-kNN, an inherently interpretable model that follows a simple case-based reasoning process. ProtoEEG-kNN reasons by comparing an EEG to similar EEGs from the training set and visually demonstrates its reasoning both in terms of IED morphology (shape) and spatial distribution (location). We show that ProtoEEG-kNN can achieve state-of-the-art accuracy in IED detection while providing explanations that experts prefer over existing approaches.

Figures (6)

• Figure 1: Top: ProtoEEG-kNN reasoning. The topographic map ("topoplot") highlights important channels as calculated by the channel-wise weights ($w_c(x)$), which are also shown in bars to the left of the input channels. From left to right, we show the input sample, the best two matches selected by our model, and the best matches chosen by each of three ablated models. Middle: ProtoEEG-kNN architecture. An input is passed through the backbone $f$ to produce a embedding. The Global Comparison Layer $\bar{g}$ computes the similarity between the embedding and each sample in the training set. The final prediction produced by $\bar{h}$ is the average label from the top-k most similar neighbors. Bottom: channel-wise similarity. The similarity along each channel is weighted by $w_c(\mathbf{x})$.
• Figure 2: Top: User Study Results. Bootstrapping with 1,000 iterations was used to calculate the mean and 95% confidence interval for Plackett-Luce weights. Bottom: PaCMAP visualization of the test set comparison spaces of ProtoPNet-kNN (left) and kNN over ISFs (right). Neighborhoods in high-dimensional space are preserved in two-dimensional PaCMAPs.
• Figure 3: Example of A User Study Ranking. Experts are asked to rank the similarity of each candidate to the input samples for 100 different inputs. The survey was conducted on a secure Qualtrics platform.
• Figure 4: Examples of Local Analysis
• Figure 5: More Examples of Local Analysis