PaperlayerView exact revision on arXiv

Source-linked AI summary

From Visual Attribution to Clinical Reasoning: Explainable Parkinson's Disease Screening from Hand-Drawn Patterns

Aritra Dey, Utsav Kumar Nareti, Chandranath Adak, Soumi Chattopadhyay, Krishna Gopal Sasmal, Saeed Anwar

arXiv:2609.14441v1cs.CVcs.AI

TL;DR

Early PD can be difficult to diagnose, and visual explanations for handwriting-based screening may not convey clinically meaningful motor abnormalities. This paper links visual attribution to structured motor descriptors and language-generated clinical rationales, reporting competitive prediction and clinically consistent explanations.

  • Problem

    Early PD can be difficult to diagnose, and visual attribution methods may not translate handwriting evidence into clinically meaningful motor explanations.

    Source: §1 · p. 2

  • Method

    The framework maps salient handwriting regions to structured motor descriptors, then uses a language-based module to generate clinical rationales.

    Source: §1 · pp. 2–3

  • Results

    The framework achieves competitive predictive performance while providing clinically consistent and interpretable explanations for hand-drawn-pattern PD screening.

    Source: §6 · p. 14

  • Takeaways & Limitations

    The approach connects visual evidence and motor descriptors to clinical reasoning in handwriting-based neurological screening.

    Source: §6 · p. 14

    Evaluation uses subject-disjoint splits, with each participant assigned exclusively to training or testing.

    Source: §3.1 · p. 4

Abstract

from arXiv · show

Parkinson's disease (PD) manifests early neuromotor impairments that become observable in controlled hand-drawn patterns such as spirals and meanders, where tremor-induced oscillations, stroke irregularity, and curvature instability reflect underlying motor degradation. In this work, we present an explainable framework for PD screening from offline hand-drawn patterns that integrates discriminative visual modeling with clinically grounded reasoning. The predictive model captures distributed structural distortions and fine-grained texture variations. It is evaluated under subject-disjoint protocols to ensure reliable generalization. To move beyond black-box classification, we introduce a multi-stage explainability pipeline that combines visual attribution with structured symptom abstraction. Salient regions are identified using attention- and gradient-based localization, followed by extraction of clinically meaningful motor descriptors quantifying contour roughness, curvature irregularity, stroke variability, and tremor-frequency energy. These descriptors are subsequently translated into coherent clinical rationales through a language-based reasoning module, linking model evidence to established PD symptomatology. By bridging visual attribution and clinical interpretation, the proposed framework advances interpretable document intelligence for neurological screening using hand-drawn patterns. Experimental results on publicly available Parkinson's disease handwriting datasets demonstrate competitive predictive performance and clinically consistent explanations.

1 Introduction

Early Parkinson’s disease can be difficult to diagnose, while drawing abnormalities offer observable indicators of impaired motor control. The paper proposes a framework that connects visual evidence from hand-drawn patterns to structured motor descriptors and clinical explanations.

1 Introduction

  • Early-stage PD may be difficult to diagnose because subtle neuromotor impairments can precede definitive clinical confirmation.Drawing abnormalities such as tremor-induced oscillations, stroke irregularity, and curvature instability are presented as indicators of impaired motor planning and execution.Source: §1 · pp. 1–2
  • Existing saliency maps and attention visualizations locate relevant image regions but do not explain the underlying motor abnormalities in clinically meaningful terms.The paper frames translating visual evidence into measurable motor characteristics as an open challenge in document-based medical AI.Source: §1 · p. 2
  • The framework links attribution maps to quantifiable motor descriptors and then to clinically grounded explanations for hand-drawn spiral and meander patterns.The descriptors include contour roughness, curvature irregularity, stroke variability, and tremor-frequency energy.Source: §1 · pp. 2–3
  • The work evaluates the framework on publicly available PD handwriting datasets using subject-disjoint protocols and reports competitive prediction with clinically meaningful explanations.The introduction also reports expert assessment and multi-rater reliability analysis of explanation quality.Source: §1 · p. 3

2 Related Work

Prior PD handwriting systems progressed from handcrafted features and classical classifiers to CNNs and transformers, but explanations often remain visual rather than clinically grounded. The paper positions its contribution as integrating attribution-guided motor descriptors with language-based clinical reasoning.

2 Related Work

  • Earlier PD handwriting methods used handcrafted features and classical classifiers, followed by CNNs and transformer-based models that learn visual representations directly.The reviewed approaches include computer vision features, fine-tuned CNNs, handwriting-dynamics models, and Vision Transformer transfer learning.Source: §2 · p. 3
  • Common post-hoc XAI methods identify image regions or feature importance but often lack semantic grounding in clinically meaningful biomarkers.The passage notes that clinicians need explanations in terms of observable motor abnormalities rather than isolated pixel-level evidence.Source: §2 · p. 3
  • The paper identifies attribution-guided motor descriptor extraction followed by language-driven clinical reasoning as an integration not explicitly addressed by prior PD screening studies.It unifies visual attribution, descriptor quantification, and semantic explanation in one screening pipeline.Source: §2 · p. 4

3 Employed Datasets

The paper evaluates HandPD and NewHandPD under a subject-disjoint protocol and assesses generated explanations with clinical and handwriting experts. Expert ratings and agreement statistics are used to characterize the explanations’ clinical consistency and structural relevance.

3.1 Hand-Drawn Pattern Datasets

  • HandPD and NewHandPD provide publicly available spiral and meander drawings collected under standardized acquisition protocols for PD screening.NewHandPD adds subjects and samples, with greater inter-subject variability and a more balanced class distribution than HandPD.Source: §3.1 · p. 4
  • Subject-disjoint splits keep each participant’s samples exclusively in training or testing to prevent identity leakage.The protocol is intended to encourage learning disease-related motor patterns rather than subject-specific drawing characteristics.Source: §3.1 · p. 4

3.2 Expert-based Evaluation of Clinical Explanations

  • The explanation assessment used a randomly selected subset of test samples from both datasets, presented with the drawing, predicted label, and generated explanation.The panel included two neurology practitioners and four handwriting-analysis experts.Source: §3.2 · pp. 4–5
  • Experts rated clinical consistency, structural relevance, and clarity on a three-point ordinal scale.Structural relevance concerned correspondence between the explanation and visible drawing distortions.Source: §3.2 · p. 5
  • Across all evaluation criteria, Fleiss’ kappa was 0.68, indicating substantial inter-rater agreement, while Krippendorff’s alpha was 0.71, confirming reliable consensus among experts under the ordinal rating scheme.Source: §3.2 · p. 5
  • The majority of generated explanations were rated clinically consistent and structurally relevant, aligning the visual highlights and descriptor-based interpretations with expert understanding of PD motor abnormalities.The paper presents this outcome as support for the structured symptom abstraction layer.Source: §3.2 · p. 5

4 Methodology

The framework classifies offline handwriting and builds explanations through a sequence from visual evidence to structured motor descriptors and clinical rationale. Its methodology combines constrained preprocessing, transformer-based visual modeling, a vision-language bridge, and jointly optimized prediction and explanation generation.

4.1 Problem Formulation

  • The framework maps an image to a PD prediction, attribution map, motor-descriptor vector, and symptom-level clinical explanation.Descriptors include contour roughness, curvature irregularity, stroke variability, and tremor-related frequency characteristics.Source: §4.1 · p. 6

4.2 Preprocessing and Augmentation

  • Images are resized and normalized, then augmented with small rotations, translations, and mild scaling while preserving pathology-relevant stroke patterns.The stated augmentation ranges rotations to ±5° and aims to support subject-disjoint evaluation without distorting tremor signatures or curvature fluctuations.Source: §4.2 · p. 6

4.3 Motor Descriptors (D)

  • The method extracts contour roughness from contour geometry and curvature irregularity from changes in local direction along the trace.The curvature score is described as capturing curvature instability and oscillatory deviations associated with tremor.Source: §4.3 · p. 7
  • Stroke intensity variability is computed from trace-pixel intensities, while a frequency score summarizes spectral energy in selected radial bands.Together with roughness and curvature irregularity, these measures form a structured descriptor vector for clinically relevant motor abnormalities.Source: §4.3 · p. 8

4.4 Vision Transformer-based Modeling

  • The vision transformer divides each handwriting image into patches, projects them with positional embeddings, and produces contextualized patch representations.The patch representations are distinguished from the CLS representation used by the classification head.Source: §4.4 · p. 8

4.5 Vision-to-Language Bridge and Explanation Generation

  • A vision-language bridge uses cross-attention between projected visual features and learnable queries to produce visual tokens for language-model input.GPT-2 is trained autoregressively using synthetic captions generated from the motor descriptors.Source: §4.5 · pp. 8–9

4.6 Classification Head and Joint Optimization

  • The classifier predicts PD from the visual encoder’s CLS representation, and joint optimization combines weighted binary cross-entropy with the language-model objective.The parameter λ controls the contribution of explanation generation to the combined objective.Source: §4.6 · p. 9

5 Results and Discussion

The framework achieved competitive screening accuracy across HandPD and NewHandPD and produced explanations aligned with structured motor descriptors. Results also show stronger performance on spirals than meanders and improved predictive and explanation measures in the full ablation model.

5.1 Experimental Setup

  • The evaluation used subject-disjoint 5-fold validation on HandPD and NewHandPD, keeping each participant’s drawings out of either both training and test sets.Reported results are test-set accuracy.Source: §5.1 · p. 9

5.2 Quantitative Comparison with State-of-the-Art

  • EfficientNet-v2-l reached 95.94% accuracy on HandPD spirals, while ViT-l-16 reached 85.13% and 85.18% on HandPD and NewHandPD meanders.The authors describe transformer performance as more stable across pattern types, while spiral accuracy exceeded meander accuracy across nearly all models.Source: §5.2 · p. 10, Table 2 · p. 10
  • 94.59% and 90.54% accuracy on HandPD spiral and meander patterns, and 92.45% and 90.57% on NewHandPD, position the proposed model competitively among prior methods.The HandPD spiral result approaches the upper range of reported deep-learning methods; NewHandPD performance is described as stable across pattern types.Source: §5.2 · p. 10, Table 3 · p. 10

5.3 Clinical Explanation and Qualitative Analysis

  • True-positive explanations aligned tremor energy and curvature variability with irregular oscillations, while true-negative explanations reflected smooth, stable strokes and controlled execution.False positives involved moderate irregularities resembling tremor; false negatives tended to show visually mild or ambiguous impairment.Source: §5.3 · p. 11, Table 4 · p. 12

5.4 Ablation Study

  • The full model improved Accuracy from 94.12% to 94.59%, F1-score from 93.71% to 94.10%, and AUC from 0.961 to 0.966 across ablation variants.The variants progressively add visual attribution, structured symptom abstraction, and language-based clinical reasoning.Source: §5.4 · p. 12, Table 5 · p. 13
  • Mean expert rating increased from 1.21 to 1.68, Fleiss’ κF from 0.61 to 0.68, and Krippendorff’s α from 0.64 to 0.71 in the full model.The table caption defines the expert rating as the mean across clinical consistency, structural relevance, and clarity.Source: §5.4 · pp. 12–13, Table 5 · p. 13

5.5 Discussions

  • Paired-prediction McNemar tests were significant across datasets (p < 0.05), while tremor energy and curvature irregularity correlated positively with predicted PD probability (r = 0.52 and r = 0.48).The authors interpret the descriptor associations as alignment between predictions and clinically interpretable motor biomarkers.Source: §5.5 · p. 13
  • Accuracy drops were below 1.5% under tested image perturbations, while borderline errors clustered around subtle irregularities and meander turns that resemble tremor distortions.Spiral attention was more concentrated along tremor-affected curves; meander attention was more spatially distributed.Source: §5.5 · pp. 13–14

6 Conclusion

The paper presents an explainability-driven framework linking visual attribution, structured motor-biomarker extraction, and language-based clinical reasoning for Parkinson’s screening from hand-drawn patterns. It reports competitive performance with clinically consistent explanations and identifies severity estimation, longitudinal analysis, and larger-dataset validation as future work.

6 Conclusion

  • The framework connects visual attribution to structured motor-biomarker extraction and language-based clinical reasoning for Parkinson’s screening from hand-drawn patterns.The conclusion frames the approach as a bridge from visual evidence to medically coherent explanations.Source: §6 · p. 14
  • Experiments show competitive predictive performance alongside clinically consistent, interpretable explanations linking model decisions to motor impairments associated with Parkinson’s disease.Source: §6 · p. 14
  • Future work will address disease-severity estimation, longitudinal analysis, and validation on larger datasets.Source: §6 · p. 14

Related papers

Loading 2609.14441v1…