An Interpretable Vision Transformer as a Fingerprint-Based Diagnostic Aid for Kabuki and Wiedemann-Steiner Syndromes

Kabuki syndrome (KS) and Wiedemann-Steiner syndrome (WSS) are rare but distinct developmental disorders that share overlapping clinical features, including neurodevelopmental delay, growth restriction, and persistent fetal fingertip pads. While genetic testing remains the diagnostic gold standard, many individuals with KS or WSS remain undiagnosed due to barriers in access to both genetic testing and expertise. Dermatoglyphic anomalies, despite being established hallmarks of several genetic syndromes, remain an underutilized diagnostic signal in the era of molecular testing. This study presents a vision transformer-based deep learning model that leverages fingerprint images to distinguish individuals with KS and WSS from unaffected controls and from one another. We evaluate model performance across three binary classification tasks. Across the three classification tasks, the model achieved AUC scores of 0.80 (control vs. KS), 0.73 (control vs. WSS), and 0.85 (KS vs. WSS), with corresponding F1 scores of 0.71, 0.72, and 0.83, respectively. Beyond classification, we apply attention-based visualizations to identify fingerprint regions most salient to model predictions, enhancing interpretability. Together, these findings suggest the presence of syndrome-specific fingerprint features, demonstrating the feasibility of a fingerprint-based artificial intelligence (AI) tool as a noninvasive, interpretable, and accessible future diagnostic aid for the early diagnosis of underdiagnosed genetic syndromes.

💡 Research Summary

This paper investigates whether fingerprint images can serve as a non‑invasive, accessible diagnostic aid for two rare developmental disorders—Kabuki syndrome (KS) and Wiedemann‑Steiner syndrome (WSS). Both conditions share overlapping clinical features such as neurodevelopmental delay, growth restriction, and persistent fetal fingertip pads, yet genetic testing, the gold standard, is often unavailable in many regions. The authors collected a total of 2,330 fingerprint images from 75 individuals with KS, 38 with WSS, and 120 unaffected controls across five sites between June 2020 and November 2024. Images were captured using a custom Android app (GenePrint) that guides users through acquisition with an external HID DigitalPersona 4500 optical scanner. After conversion to 8‑bit RGB PNG, images were inverted and enhanced with a Gabor‑filter pipeline to reduce noise. Low‑quality scans (NFIQ‑2 score < 2) were discarded, leaving 2,109 high‑quality images for analysis.

The classification model is a Vision Transformer (ViT). Input images are resized to 224 × 224 px, split into non‑overlapping 16 × 16 patches, and linearly projected to embeddings of dimension 512 for control‑vs‑KS and control‑vs‑WSS tasks, or 256 for KS‑vs‑WSS. A learnable class token and fixed positional encodings are added, followed by three transformer encoder blocks, each containing four self‑attention heads and a feed‑forward layer (hidden dimension 1,024 for the first two tasks, 512 for KS‑vs‑WSS). The class token output feeds a linear classification head. To mitigate overfitting on the limited data, five independent models are trained for 10 epochs each (Adam optimizer, learning rate 3 × 10⁻⁴, no dropout) and their logits are averaged to produce ensemble predictions.

Performance is evaluated on a held‑out test set (80 % training / 20 % testing split at the participant level). The three binary tasks achieve the following results:

• Control vs. KS – Accuracy 0.72, Precision 0.71, Recall 0.70, F1 0.71, AUC 0.80.
• Control vs. WSS – Accuracy 0.80, Precision 0.73, Recall 0.73, F1 0.72, AUC 0.73.
• KS vs. WSS – Accuracy 0.88, Precision 0.84, Recall 0.82, F1 0.83, AUC 0.85.

Beyond raw metrics, the authors extract self‑attention weights from the class token, average them across all layers and heads, reshape to the patch grid, normalize, and up‑sample to the original image resolution. Overlaying these attention heatmaps on the grayscale fingerprints reveals which regions most influence the model’s decision. In control‑vs‑KS and control‑vs‑WSS, the central fingertip area—where persistent fetal pads are located—receives the strongest attention, but additional salient regions include ridge curvature, bifurcations, and overall pattern geometry. In the KS‑vs‑WSS comparison, attention spreads across broader fingerprint features, suggesting that subtle, syndrome‑specific dermatoglyphic signatures exist beyond the obvious pad persistence.

The study acknowledges several limitations: the sample size, especially for WSS, is modest; the workflow relies on a dedicated optical scanner rather than a smartphone camera, which may limit scalability; and the biological interpretation of the learned features (e.g., correlation with specific KMT2A/KMT2D mutations) remains unexplored. Future work is proposed to (1) validate a purely smartphone‑based acquisition pipeline, (2) expand the dataset through multi‑center collaborations, (3) integrate multimodal data (facial images, speech, clinical notes), and (4) investigate genotype‑phenotype links between genetic variants and fingerprint patterns.

In conclusion, the paper demonstrates that a Vision Transformer can extract discriminative, interpretable features from fingerprint images to differentiate KS, WSS, and healthy controls. The attention‑based visualizations provide a transparent view of the model’s reasoning, supporting the notion that dermatoglyphic anomalies contain syndrome‑specific information. This proof‑of‑concept suggests that AI‑driven fingerprint analysis could become a rapid, low‑cost, and widely deployable screening tool for rare genetic disorders, especially in settings where conventional genetic testing is inaccessible.