Research prototypeEEG · Edge AI
A wearable-to-edge concept for distinguishing ictal, preictal, postictal, and normal EEG states and exploring whether a useful warning can be produced before a seizure.
- Data and signals
- TUH EEG records, reduced to an eight-channel configuration.
- Pipeline
- Multi-scale temporal features, channel attention and fusion, temporal context modeling, OpenBCI Cyton, Raspberry Pi, and a Flutter companion app.
- Intended workflow
- Wearable sensing, edge inference, alerts, emergency location sharing, and medication tracking.
Evidence boundary: the local brief sets a two-to-five-minute pre-seizure warning as a design target; it does not provide prospective clinical validation for that target.
Competition projectChest X-ray · Explainability
A frontal chest X-ray system that combines anatomical segmentation, cardiothoracic measurements, CNN evidence, and tabular classification instead of relying on a single opaque score.
- Data
- Project materials describe PadChest, MIMIC-CXR, CheXpert, VinDr-CXR, NIH ChestX-ray14, and CheXmask-derived masks.
- Models
- U-Net++ for the heart, DeepLabV3+ with ResNet101 for the lungs, and LightGBM over CTR, area features, and CNN probability.
- Reported results
- Heart IoU 87.26%; right/left lung IoU 91.11%/89.53%; classifier F1 91.4%; CTR SMAPE 7.5%.
Interpretability: the report includes cardiothoracic ratio, heart/lung area features, Grad-CAM, and SHAP. It also documents StyleGAN2-ADA synthetic-data experiments.
Competition projectColonoscopy · Segmentation
A multi-dataset study of polyp masks in colonoscopy frames, built around reproducible preprocessing and comparative segmentation experiments.
- Models explored
- Attention U-Net, U-Net++, DeepLabV3+, PVT-Cascade, and SSFormer-S with a progressive refinement network.
- Ablations
- Learning-rate comparisons, bilateral filtering, elastic deformation, image generators, and cosine-annealing warm restarts.
- Reported results
- The report records U-Net++ validation mean IoU of 0.8907 and PolypDB mean IoU of 0.8655, alongside uneven results across external datasets.
Generalization note: the documented cross-dataset variance is retained because aggregate validation performance alone can hide distribution shift.
Research conceptEchocardiography · Video
An echocardiography-video concept for segmenting the left ventricle, estimating end-diastolic and end-systolic volumes, and deriving ejection fraction.
- Visual model
- Multi-scale feature extraction with spatial attention and feature fusion.
- Temporal model
- Motion analysis across the cardiac cycle rather than isolated-frame classification.
- Planned output
- An EF estimate and low-versus-preserved EF decision-support category.
Evidence boundary: the available brief describes the architecture and intended outputs but contains no evaluation cohort or validated performance result.
Computational studyBioinformatics · Docking
Two linked computational workstreams: multimodal mutation analysis over sequence and structure, followed by a focused study of how the R166H variant changes G6PD structure and small-molecule binding profiles.
- Mutation models
- NCBI cDNA and protein data, AlphaFold structures, a GNN, sequence-based XGBoost/ensemble models, and a combined multimodal model. The brief reports F1 scores of 72%, 84%, and 94% respectively.
- G6PD case study
- Wild-type and R166H G6PD sequences modeled with AlphaFold and compared at the structural level.
- Screen
- More than 8,000 DrugCentral compounds were filtered with RDKit; 3,799 molecules were docked against both forms with GNINA.
- Analysis
- Binding energies, mutant-versus-wild-type ΔΔG, CNN pose scores, distributions, and top-ligand comparisons.
Evidence boundary: the F1 values are project-reported and the docking results are computational hypotheses. Neither establishes a therapy or proves restoration of protein function without independent and laboratory validation.
Exploratory studyGCK · Variant classification
A small-data experiment classifying GCK variants associated with MODY2 from local sequence context.
- Dataset
- 317 pathogenic and 39 benign variants assembled from ClinVar and LOVD.
- Method
- 3-mer encoding, SMOTE within the training workflow, a 20% held test split, and a voting ensemble spanning tree, linear, kernel, and boosting models.
- Reported results
- Accuracy 78.35%, ROC AUC 0.7987, and PR AUC 0.8313.
Evidence boundary: 356 labeled variants are not sufficient for a clinical claim; the result is presented as an exploratory baseline.
Prototype conceptFundus · OCT
A multimodal ophthalmic-imaging concept for disease classification and lesion localization across fundus photographs and OCT scans.
- Scope
- The brief targets more than 25 eye conditions across the two modalities.
- Preprocessing
- Contrast adjustment, Gaussian smoothing, and histogram equalization.
- Model direction
- CNNs, attention, transfer learning, and a prototype Gemini-connected interpretation layer.
Evidence boundary: accuracy, F1, and IoU are proposed evaluation metrics; the available document does not report a completed validation study.
Prototype conceptHeart sounds · Signal processing
A digital-auscultation workflow for cleaning heart sounds, extracting acoustic features, and supporting normal-versus-murmur screening.
- Features
- MFCC, RMS energy, zero-crossing rate, and normalized temporal representations.
- Models considered
- Temporal CNN/RNN approaches and classical SVM or random-forest baselines.
- Interface
- Mobile or web software paired with a Bluetooth-capable digital stethoscope.
Evidence boundary: the project file contains a system design but no reported patient-level evaluation.
Early conceptEMG · Neuromuscular
An early research direction for using electromyography signals to distinguish neuropathic and myopathic patterns.
- Available scope
- The local archive currently contains only the problem statement.
- Required next evidence
- A defined dataset, patient-level split, signal preprocessing protocol, baseline models, error analysis, and clinical evaluation plan.
Status: listed for completeness as an early concept; no implementation or result is claimed.
