Segment-Level Risk Discovery in Online Handwriting for Alzheimer's Disease Detection

I. INTRODUCTION

• Online handwriting can capture fine-grained cognitive-motor behavior, but global representations may be affected by writing style, task variation, and acquisition noise while abnormalities may be localized.The paper motivates segment-level discovery as a way to focus on disease-relevant regions.Source: §I · pp. 1–2

III. PROPOSED APPROACH

• NormPaST-Risk uses weakly supervised segment-risk learning because trajectory-level labels lack annotations identifying abnormal local regions.It combines multi-scale temporal encoding, state-consistent Paper/Air segmentation, selective state-space modeling, healthy normative learning, and task-aware multi-expert risk estimation.Source: §I · p. 2, §III · p. 3

I. INTRODUCTION

• The framework estimates segment-level disease relevance to adaptively weight local representations while retaining the full trajectory context.The introduction also reports that experiments outperform representative machine-learning and sequence-modeling baselines, with ablation and segment-deletion studies supporting component contributions and the discriminative value of discovered segments.Source: §I · p. 2

II. RELATED WORK

• Prior work progressed from handcrafted task-level features to learned sequential, attention-based, multimodal, and stroke-level representations, but often predicts from tasks or samples rather than learning disease-relevant segments.Existing fine-grained approaches also rely mainly on predefined stroke units or post-hoc interpretation.Source: §II · pp. 2–3

A. Input Representation

• The input representation defines the raw trajectory and valid, paper, and air masks for its points.Paper and air states are mutually exclusive for each valid point.Source: A. Input Representation · p. 3

B. Multi-Scale Trajectory Encoder

• The multi-scale encoder applies temporal streams with different kernel sizes and dilation rates while preserving the original sequence length.Each stream encodes the projected, positionally augmented trajectory.Source: B. Multi-Scale Trajectory Encoder · p. 3
• Learned scale weights adaptively fuse the temporal streams at each raw point into a point-wise sequence.The fused representation combines information across temporal scales.Source: B. Multi-Scale Trajectory Encoder · p. 3

C. Temporal Segment Aggregation

• State-consistent temporal segmentation avoids inefficient, noise-sensitive point-level encoding and prevents downsampling from merging pen-down and pen-up states.Each Paper/Air transition ends an interval, so all points within a segment share one pen state.Source: C. Temporal Segment Aggregation · pp. 3–4
• The aggregator partitions valid points into intervals and mean-pools both the raw trajectory and fused point-level features over the same layout.Each resulting token remains aligned with a raw-point interval and a single Paper/Air state.Source: C. Temporal Segment Aggregation · pp. 3–4

D. Length-Aware Segment Token Fusion

• Because mean pooling can dilute salient responses, the model adds channel-wise max pooling and combines mean and max tokens with a learnable convex mixture.The mixture retains the stable mean-pooling path while providing an additional length-independent path for salient responses.Source: D. Length-Aware Segment Token Fusion · p. 4
• Segment duration is encoded with two descriptors and incorporated into the final segment token.Segments may have different true lengths, including when a Paper/Air transition terminates one early.Source: D. Length-Aware Segment Token Fusion · p. 4

E. Selective Paper-Air State-Space Encoder

• Paper/Air state modulates intermediate trajectory representations at each layer before selective state-space dynamics model long-range dependencies.On-paper movements primarily reflect executed motor behavior, while in-air movements may encode planning, hesitation, and cognitive-motor transitions.Source: E. Selective Paper-Air State-Space… · p. 5
• Each layer learns its own representation and modulation effect for the observed Paper/Air state, which remains unchanged across layers.The state-specific gate is predicted from the preceding representation and a layer-specific latent embedding.Source: E. Selective Paper-Air State-Space… · p. 5
• A normalized hidden-state change score calibrates learned segment risk rather than acting as an independent disease indicator.The score is computed from changes between consecutive context-aware states.Source: E. Selective Paper-Air State-Space… · p. 5

F. Healthy Normative Branch

• The healthy normative branch predicts the subsequent aggregated trajectory token from each context-aware segment state.The prediction objective is optimized exclusively on healthy-control samples to capture healthy handwriting dynamics.Source: F. Healthy Normative Branch · p. 6
• For consecutive valid segments, normative deviation measures the difference between the predicted and observed trajectory tokens and is aligned with segment tokens.The first segment's deviation is set to zero; F denotes trajectory-feature dimensionality.Source: F. Healthy Normative Branch · p. 6
• At inference, deviations from healthy regularities provide auxiliary cues for segment-risk calibration and subject-level classification.The sample-level normative score is computed from unnormalized deviation.Source: F. Healthy Normative Branch · p. 6

G. Task-Aware Multi-Expert Segment-Risk Modeling

• Each expert produces a segment-risk logit, and task identity informs routing across experts.The router combines a global state formed by masked mean pooling with a task embedding, then converts expert logits into normalized weights.Source: G. Task-Aware Multi-Expert… · p. 6
• The final segment-risk logit is calibrated by hidden-state change and healthy-normative deviation.The normative deviation is detached in this calibration path so AD labels do not directly optimize the healthy-only predictor through the segment-risk objective.Source: G. Task-Aware Multi-Expert… · p. 6
• Temporal risk weights are used to pool segment information into a handwriting-level risk representation.The section describes a segment-level AD probability and masked-softmax temporal risk weights.Source: G. Task-Aware Multi-Expert… · pp. 6–7

H. Final Classification

• The projected sample-level normative score is added to segment risk in the final classification stage.A learnable projection maps the score into the same feature space as segment risk.Source: H. Final Classification · p. 7
• The normative score is detached before addition so the classification objective does not directly optimize the healthy normative predictor.Source: H. Final Classification · p. 7
• The section then forms the AD/HC prediction from the resulting risk representation.Source: H. Final Classification · p. 7

• The segment-level and normative loss weights are 0.1 and 0.2, respectively, to keep auxiliary objectives effective without overwhelming classification.The paper describes this weighting as supporting balanced and stable optimization.Source: §I · p. 7
• The sample-level classification loss uses standard cross-entropy.Source: §I · p. 7
• Weakly supervised segment-level multiple-instance learning uses trajectory labels to encourage high-risk segments in AD samples and suppress strong AD-like responses in healthy controls.Source: §I · p. 7
• The normative loss is applied only to healthy controls; it is zero when a mini-batch contains no healthy-control sample.AD samples do not contribute, so the predictor learns healthy segment-level trajectory regularities rather than disease-specific patterns.Source: §I · p. 7

A. Dataset

• DARWIN-RAW includes 174 participants (89 AD, 85 HC), each completing 25 memory, copying, and graphic handwriting tasks recorded at 200 Hz.Source: A. Dataset · p. 7

B. Experimental Setup

• Five-fold evaluation kept each participant's samples in one split and compared baselines under matched subject-level splits, input features, and evaluation protocol.Source: B. Experimental Setup · p. 7
• Balanced accuracy was primary, with Precision-AD, F1-AD, sensitivity, specificity, and AUC also reported.Source: B. Experimental Setup · p. 7

C. Comparison with Representative Baselines

• Mamba's 97.71% sensitivity paired with 79.93% specificity, versus NormPaST-Risk's 93.33% sensitivity and 91.77% specificity, illustrating that peak sensitivity need not mean balanced performance.Source: C. Comparison with Representative… · p. 8
• NormPaST-Risk combines multi-scale dynamics, Paper–Air state-aware modeling, long-range dependencies, healthy-normative deviations, and segment-risk calibration to emphasize locally informative abnormal regions.Source: C. Comparison with Representative… · pp. 8–9
• 92.55% balanced accuracy, 95.84% AUC, 92.43% Precision-AD, 92.74% F1-AD, and 91.77% specificity were best among methods on five of six metrics.Source: C. Comparison with Representative… · p. 8, Table II · p. 9

D. Ablation Study

• Removing the segment-risk module produced the largest degradation: balanced accuracy fell from 92.55% to 88.56%, F1-AD from 92.74% to 88.46%, and sensitivity from 93.33% to 87.71%.The authors report that identifying and aggregating disease-relevant local segments is more effective than treating all trajectory regions equally.Source: D. Ablation Study · p. 9
• Removing Paper-Air modulation reduced balanced accuracy to 89.67% and sensitivity to 88.76%; removing the Selective SSM reduced sensitivity to 85.36%, despite specificity rising to 94.12%.The authors associate Paper-Air distinctions with useful state-dependent information and long-range state-space modeling with identifying AD participants.Source: D. Ablation Study · p. 9
• Eight experts achieved the best overall performance, while increasing top-k from 0.1 to 0.5 improved performance before a slight decline at 0.7.The authors describe diminishing returns from four to eight experts and suggest that including more lower-ranked segments can dilute discriminative evidence.Source: Figure 4 · p. 9, D. Ablation Study · p. 10
• Without multi-scale temporal modeling, balanced accuracy and F1-AD fell to 90.23% and 90.46%; removing the Healthy Normative Branch reduced them to 90.95% balanced accuracy and 89.93% sensitivity.The normative branch provides complementary information for calibrating segment-level risk through deviations from healthy handwriting dynamics.Source: D. Ablation Study · p. 9

E. Interpretability and Faithfulness Analysis

• AD participants had higher mean, maximum, and top-50% risks than HC participants, with Cohen’s d greater than 2.0 for each aggregation.The absolute group differences were 0.058, 0.096, and 0.074 for mean, maximum, and top-50% aggregation, respectively.Source: Table IV · p. 10, E. Interpretability and Faithfulness… · p. 10
• Removing 10% of highest-risk segments decreased balanced accuracy by 7.35 points and AUC by 1.96 points, versus a 1.21-point balanced-accuracy decrease and almost no AUC change for low-risk deletion.At 50% deletion, high-risk removal decreased balanced accuracy by 12.84 points and AUC by 6.22 points, while low-risk removal decreased balanced accuracy by 3.11 points with negligible AUC degradation.Source: Table V · p. 10, E. Interpretability and Faithfulness… · p. 10
• Across selected tasks, the representative AD participant showed frequent, pronounced localized risk regions, unlike the HC participant’s generally low and sparse regions.The figure contrasts original trajectories, selected highest-risk segments, and trajectory-level risk heatmaps.Source: Figure 5 · p. 11, E. Interpretability and Faithfulness… · p. 11
• High-risk regions often coincide with curved strokes, direction changes, transitions, and Paper-Air switches, but are model-derived associations rather than direct evidence of specific cognitive or motor impairments.The visualizations are qualitative model evidence, not direct clinical biomarkers.Source: E. Interpretability and Faithfulness… · pp. 11–12

V. CONCLUSION

• NormPaST-Risk combines multi-scale temporal encoding, Paper-Air and selective state-space modeling, healthy normative learning, and task-aware segment-risk estimation.Learned segment risks can be projected onto the original handwriting trajectory as interpretable evidence.Source: §V · p. 12
• NormPaST-Risk reached 92.55% BAcc and 95.84% AUC; deleting high-risk segments degraded performance more than deleting low-risk or randomly selected segments.The authors report that these findings support the discriminative value of the discovered local evidence.Source: §V · p. 12
• Fixed segments of 128 trajectory points may split strokes, letters, or words, weakening semantic continuity and the interpretability of high-risk regions.The authors suggest adaptive or stroke-aware segmentation as a future direction.Source: §V · p. 12
• Evaluation is mainly on one online handwriting dataset, leaving robustness across acquisition devices, tasks, and populations in need of validation.The normative branch may also be sensitive to the diversity and representativeness of its healthy reference population.Source: §V · p. 12
• Because segment risks are learned without manual segment annotations, the identified regions are model-derived evidence rather than direct clinical biomarkers.The authors identify clinical verification of discovered patterns as future work.Source: §V · p. 12