PointExplainer: Towards Transparent Parkinson's Disease Diagnosis

TL;DR

This work addresses the interpretability gap in using digitized hand-drawn signals for early Parkinson's disease diagnosis. It proposes PointExplainer, which encodes hand-drawn trajectories as 3D point clouds and learns a local, interpretable surrogate to approximate a black-box diagnostic model, producing discrete attributions over hand-drawn segments. The framework introduces two families of consistency measures to verify explanation faithfulness and demonstrates robust diagnostic performance and faithful explanations across SST, DST, and DraWritePD datasets, with public code available. The approach aims to enhance clinical trust and enable localization of diagnostically relevant hand-drawn regions, advancing practical adoption of digital handwriting analysis in PD care.

Abstract

Deep neural networks have shown potential in analyzing digitized hand-drawn signals for early diagnosis of Parkinson's disease. However, the lack of clear interpretability in existing diagnostic methods presents a challenge to clinical trust. In this paper, we propose PointExplainer, an explainable diagnostic strategy to identify hand-drawn regions that drive model diagnosis. Specifically, PointExplainer assigns discrete attribution values to hand-drawn segments, explicitly quantifying their relative contributions to the model's decision. Its key components include: (i) a diagnosis module, which encodes hand-drawn signals into 3D point clouds to represent hand-drawn trajectories, and (ii) an explanation module, which trains an interpretable surrogate model to approximate the local behavior of the black-box diagnostic model. We also introduce consistency measures to further address the issue of faithfulness in explanations. Extensive experiments on two benchmark datasets and a newly constructed dataset show that PointExplainer can provide intuitive explanations with no diagnostic performance degradation. The source code is available at https://github.com/chaoxuewang/PointExplainer.

Figures (7)

• Figure 1: Visualization of diagnostic results for the same instance. The 2D spiral hand-drawn trajectory embeds an additional hand-drawn feature (e.g., radius, which is the distance of the hand-drawn point relative to the template center.) as height point by point to form a point cloud. (a) Most methods only output the final model decision, whereas (b) our PointExplainer enhances interpretability by generating an attribution map. This map overlays attribution values as color intensities onto the hand-drawn trajectory, intuitively highlighting key regions. This makes the model decision more transparent. Blue represents a tendency towards healthy controls (HC), and red represents a tendency towards Parkinson’s disease (PD).
• Figure 2: Overview of the proposed PointExplainer framework, consisting of four main modules: (a) point cloud representation, (b) black-box diagnosis, (c) point perturbation explanation, and (d) explanation fidelity verification. First, the hand-drawn signal is encoded point by point into a point cloud. The black-box diagnostic model then classifies the constructed point cloud representation. Next, an interpretable model identifies key hand-drawn regions, and finally, the reliability of explanations is quantitatively verified. Further details can be found in Section \ref{['sec:method']}.
• Figure 3: Illustration of Point Perturbation. All points within the superpoint $\mathbf{S}$ are shifted toward its center, eliminating the original local spatial semantic structure.
• Figure 4: Training and performance curves. We show the training process of PointExplainer and the diagnostic performance under different decision thresholds. The curves represent the mean results from three-fold cross-validation, with shaded regions indicating the range between the minimum and maximum values. The optimal decision thresholds ($\alpha$) for the SST, DST, and DraWritePD datasets are 0.52, 0.46, and 0.50, respectively.
• Figure 5: Qualitative explanation results and the corresponding perturbation analysis. We conduct superpoint-level quantitative perturbation experiments based on the explanation results (i.e., attribution maps). For each instance, the left side presents the ground-truth label, model-predicted label, and corresponding attribution map. The right side shows how the model decision changes when each superpoint is perturbed individually (see Section \ref{['sec:perturnation']} for details on the perturbation strategy). In the attribution map, blue indicates attribution values less than 0 (i.e., the superpoint contributes negatively, favoring HC), and red indicates attribution values greater than 0 (i.e., the superpoint contributes positively, favoring PD). In the bar chart, the x-axis represents the index of superpoints reordered by their attribution values, with the corresponding attribution value labeled on top of each bar, while the y-axis represents the relative percentage change in the predicted probability, computed according to Eq. $(F(\mathbf{X})-F(\mathbf{X} \backslash \mathbf{S}_{j_{s}}))/F(\mathbf{X})$.