Research

Parkinson's, multiple sclerosis and early dementia share a diagnostic problem: the earliest signal is a change in motion or lesion pattern too subtle for a single visit to catch. HCF's work here builds models that read that change directly from cheap, repeatable inputs — handwriting kinematics and spatio-temporal graphs for Parkinsonian tremor, time-aware lesion-saliency transformers for MS progression, and a hybrid generative-heuristic pipeline combining Tri-GAN synthesis with metaheuristic neural architecture search for early dementia detection.

The Parkinson's work, HybridParkinsonModelV4.1, fuses an adaptive spatio-temporal graph convolutional network with discrete wavelet transform features and clinical handwriting markers, evaluated on PaHaW with subject-independent five-fold GroupKFold to avoid the leakage that comes from splitting by sample instead of by patient — a discipline that runs through the whole portfolio, below.

This line asks a harder question of every model than accuracy alone: does it generalise, and does it survive the edge device it has to run on? The ECG work on PTBDB uses two sub-8,000-parameter architectures, CompactCNN1D and TinyTCN, built for INT8 quantisation and audited for out-of-distribution generalisation, near-duplicate leakage and six-method explainability — because a benchmark score that survives a leakage audit is worth more than one that doesn't.

The same discipline produced the portfolio's most consequential negative result: a quantum kernel SVM for wearable EDA stress detection was tested against a classical MLP autoencoder on WESAD and lost decisively — 0.601 AUC against 0.832, across 15 subjects and 921 windows. That result is reported here, not around, alongside a separate contrastive TCN approach for physiological stress signals too muted for standard classifiers to separate.

PCOS affects an estimated one in five women in India and is routinely underdiagnosed because its markers are distributed across symptoms nobody connects at a single visit. HCF's published PCOS detection model addresses that by learning across the combined marker set rather than any one test in isolation.

Two review papers extend this line into the molecular biology underneath metabolic and inflammatory disease: mitochondrial dysfunction as a common mechanism across chronic conditions, and the neuropeptide pathways behind diabetic wound healing and gangrene formation. Both are literature syntheses rather than novel models, written to give the applied work above a documented mechanistic foundation rather than treating symptoms as unconnected labels.

A model that flags a lesion without saying why is not useful to a clinician who has to act on it, and a model whose explanation depends on the dataset it happened to be trained on is not trustworthy at all. The imaging line treats explainability as a claim to be tested, not a feature to be added: mask-supervised attention for dermoscopic classification is evaluated in a controlled cross-dataset setting specifically to check whether its saliency maps hold up outside their training distribution.

This is the same standard applied across the portfolio's other imaging and signal work — SHAP and Grad-CAM are run as verification tools against every diagnostic model that reaches review, not as a figure generated once for the paper.

A screening model is only as trustworthy as the pipeline that feeds it, and in a connected clinic that pipeline is a set of IoMT devices and a network nobody fully controls. Med-ZT and the companion context-aware zero-trust framework for edge IoMT environments treat security and explainability as one problem: every access decision has to be both correct and auditable by someone who is not a security engineer.

Federated learning is the same principle applied to the model rather than the network — QS-FedHealth trains a quantum-sparse federated model for ECG classification on PTB-XL without centralising patient data, and a systematic review of federated learning for IoMT security and privacy sets out where that approach currently holds up and where it still doesn't.