Diabetic foot ulcers (DFUs) are a major cause of lower-extremity amputation, and image-based decision support may assist screening and preliminary triage where specialist access is limited. We propose DFU-MambaKAN, a hybrid architecture combining a pure-PyTorch selective state-space (Mamba) block with a Gaussian radial-basis-function Kolmogorov–Arnold Network (RBF-KAN) feed-forward layer. The study evaluates two distinct tasks: (1) binary normal-versus-ulcer screening and (2) a dataset-specific four-class Wagner-Meggitt severity-grading task. DFU-MambaKAN contains 1.062 million parameters and was compared with ResNet50, EfficientNet-B0, MobileNetV3-Small, and ViT-Tiny under the manuscript’s common downstream training protocol. The reported values are point estimates from one deterministic 70/15/15 split generated with seed 42; no confidence intervals, repeated-seed averages, or inferential significance tests were obtained. On binary screening, DFU-MambaKAN achieved 97.17% accuracy, macro-F1 0.970, and AUC 0.994. On four-class grading, it achieved 67.75% accuracy, macro-F1 0.677, and AUC 0.895, whereas the baselines achieved 98.14–99.00% accuracy. This large gap means that the present evidence does not establish competitiveness for multiclass severity grading. Possible contributors include training duration and convergence, dataset size and class definitions, label or image-quality uncertainty, hyperparameter selection, global token mixing, and architecture–task mismatch; extended learning-curve analysis and multi-seed evaluation are needed to test these explanations. An automated technical audit identified 342 exact duplicates among 1055 class-folder images in Dataset A (32.4%); no deduplicated-versus-nondeduplicated performance comparison was performed, so this finding is reported as a methodological caution rather than proof of accuracy inflation. The model’s 1.062M parameter count supports parameter-efficient storage, but its measured single-image GPU latency was 90.54–93.14 ms, substantially higher than the baselines. Accordingly, the current evidence supports further investigation for batched or queued screening more strongly than immediate real-time mobile deployment. External clinical validation, multi-seed analysis, statistical testing, convergence studies, and component ablations remain necessary.

Atrial fibrillation (AF) affects more than 37 million people worldwide and substantially raises the risk of stroke, heart failure, and premature death. Wearable single-lead ECG devices offer a natural platform for continuous AF surveillance, yet automated detection from short, noisy recordings remains an unsolved clinical engineering problem. This paper presents a sys tematic benchmark of three deep learning architectures—a one dimensional ResNet-34 (1D-ResNet34), a Temporal Convolutional Network (TCN), and a novel State-Space Model (MambaECG)— evaluated under identical training protocols on two indepen dent public datasets: the PhysioNet Computing in Cardiology Challenge 2017 corpus (8,528 recordings) and the Chapman Shaoxing 12-lead ECG database (45,152 recordings, Lead I used). All models were trained exclusively on single-lead PhysioNet recordings and tested on Chapman–Shaoxing Lead I without fine-tuning, constituting the first cross-institutional generalisation study for this architectural family. The TCN achieves the highest AUROC of 0.957 with only 0.19M parameters, outperforming 1D-ResNet34 (AUROC 0.955, 35.34M parameters) and MambaECG (AUROC 0.928, 0.65M parameters). Notably, all three architectures demonstrate positive zero-shot transfer to Chapman–Shaoxing, with AUROCs of 0.961, 0.963, and 0.948 respectively—without any target-domain exposure during training. MambaECG achieves the fastest in ference at 0.04 ms per recording, enabling real-time deployment on resource-constrained wearable hardware. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis confirms that MambaECG correctly localises clinically meaningful ECG fea tures: QRS complexes and P-waves for normal rhythm, irregular RR intervals and fibrillatory baseline for AF, and premature ventricular complex morphology for Other rhythm. These results provide evidence-based architectural guidance for wearable AF screening systems and demonstrate that lightweight temporal models match or exceed deep residual networks at a fraction of the computational cost.

Atrial fibrillation (AF) detection from short single-lead electrocardiogram (ECG) recordings requires a balance between rhythm discrimination and computational efficiency. This study compares a one-dimensional ResNet-34 (1D-ResNet34), a Temporal Convolutional Network (TCN), and MambaECG, a compact Mamba-inspired gated temporal model, under the same reported training protocol on the PhysioNet Computing in Cardiology Challenge 2017 corpus. The PhysioNet dataset contains 8528 recordings. For independent dataset evaluation, the Chapman–Shaoxing database contains 45,152 recordings, of which 35,339 were used: 1780 AF records and 33,559 selected sinus-rhythm comparator records after excluding 9813 other-arrhythmia records. Models trained on PhysioNet were evaluated on Chapman Lead I without target-domain fine-tuning, forming a zero-shot evaluation on an independent but simplified AF-versus-selected-sinus task. On PhysioNet, TCN achieved the highest AUROC (0.957), F1 (0.594), sensitivity (0.883), and specificity (0.914) with 0.19 M parameters. ResNet1D34 achieved AUROC 0.955 with 35.34 M parameters, while MambaECG achieved AUROC 0.928 with 0.65 M parameters and the lowest reported GPU inference latency of 0.04 ms per recording. Chapman selected-task AUROCs were 0.948, 0.963, and 0.961 for MambaECG, ResNet1D34, and TCN, respectively. Because the PhysioNet and Chapman non-AF comparator definitions differ, these AUROCs are not treated as evidence that performance improved across domains. Grad-CAM examples are used as qualitative interpretability illustrations rather than clinical validation. The results support a controlled comparison of accuracy, parameter efficiency, and measured latency while identifying limitations in label harmonisation, padding, partition documentation, and the Mamba-inspired proxy implementation.

Diabetic foot ulcers (DFUs) are a leading cause of lower-extremity amputation world wide, yet routine assessment still relies on manual visual inspection by trained clinicians, a resource that is subjective and unevenly available in low-resource settings. We pro pose DFU-MambaKAN,alightweight hybrid architecture that combines a pure-Py Torch selective state-space (Mamba) block with a Gaussian radial-basis-function Kolmogorov Arnold Network (RBF-KAN) feed-forward layer. The model targets two clinically moti vated tasks organized as a screening-then-grading cascade: (1) binary normal-versus-ulcer screening and (2) four-class Wagner-Meggitt severity grading. It uses only 1.06 million parameters, 1.4 to 22 times fewer than four standard baselines (ResNet50, Efficient Net B0, MobileNetV3-Small, ViT-Tiny) evaluated under an identical, deduplicated, fixed-seed train/validation/test protocol. On the screening task, DFU-Mamba KAN reaches 97.17% accuracy (macro-F1 0.970, AUC 0.994), within 0.9–2.9 points of all four baselines. On the severity-grading task it reaches 67.75% accuracy (macro-F1 0.677, AUC 0.895), a much larger gap than the baselines (98.1–99.0%), which we analyze in detail and attribute mainly to under-convergence of the sequential state-space scan on a problem with ten times more data and twice as many classes, rather than to a fundamental limitation of the architec- ture. We also checked the two public Kaggle-hosted DFU datasets used here for duplicate images and found a 32.4% exact-duplicate rate in one of them; left uncorrected, this in- flates reported accuracy through train/test leakage. The main contribution of this paper is not a state-of-the-art accuracy claim. It is a reproducible, efficiency-aware benchmark, a parameter-efficient architecture, and a reported limitation, together with concrete next steps toward low-cost, edge-deployable DFU triage tools.