Work where methods
meet the field.

WORK PERFORMED

1. 01Dataset preparation, training-script adjustment and evaluation workflows.
2. 02Pretrained weights, fine-tuning, learning-rate, optimizer, scheduler, AMP, EMA and checkpoint-resume exploration.
3. 03Accuracy, F1, confusion matrices and inference / compute comparison.

The explored model routes are not presented as a claim that every model was fully trained under one identical experiment.

WORK PERFORMED

1. 01Selected-modality extraction using Chinese BERT, OpenSMILE, TalkNet ASD, OpenFace and FFmpeg.
2. 02Padding / truncation, valid-length and status recording, with partial failures kept explicit.
3. 03Machine-readable NPZ and JSON outputs plus human-reviewable HTML reports.

Missing transcripts, faces or failed modalities are recorded as unavailable evidence; the pipeline does not fabricate them.

WORK PERFORMED

1. 01Python / C++ interfaces, training, prediction, model save/load and file-based workflows.
2. 02Raw pixels, HOG + PCA, Gaussian convolution features, patches, One-vs-Rest and Bayesian ensembles.
3. 035-fold out-of-fold evaluation, mutual-information weighting and difficult-class-pair analysis.

Leakage-like Heart Disease results caused by duplicate samples were excluded rather than presented as success.