Research

Motivation: Self-supervised ECG models are usually judged by accuracy on clean test data. Clinical use also needs predictions that stay accurate, honestly confident, and faithfully explained when signals are noisy, leads are missing, or the recordings come from a different hospital.

Approach: Using the same 1-D CNN encoder, we compare supervised training with contrastive, masked, and hybrid self-supervised pretraining (unlabeled Chapman-Shaoxing/Ningbo ECGs), fine-tuned on PTB-XL with 10% and 100% of the labels. The models are stress-tested under recorded electrode-motion, muscle, and baseline-wander noise (MIT-BIH NSTDB), lead loss, and two external databases (CPSC2018, Georgia), measuring accuracy, calibration (class-wise ECE, temperature scaling), and Integrated-Gradients explanation faithfulness and stability, with patient-level bootstrap statistics.

Key findings:

• Masked pretraining generalized best to unseen hospitals, especially with 10% of the labels, but no self-supervised objective improved robustness to noise or missing leads
• All models were well calibrated in distribution, but calibration error grew several-fold under dataset shift and lead loss, and temperature scaling did not repair it
• The models most accurate under dataset shift had the least stable explanations: no objective was best on every reliability axis
• Fine-tuning on corrupted inputs removed most of the lead-loss failure without reducing clean accuracy

Status: M.Sc. thesis research — manuscript in preparation for journal submission

Motivation: Potato is a staple crop in Bangladesh, and leaf diseases cause significant yield losses. Early, accurate detection enables timely intervention, but expert pathologists are scarce in rural areas.

Approach: We build a deep learning pipeline for automated leaf disease classification from smartphone images, paired with an actionable guidance system that recommends treatment protocols to farmers in local language.

Key contributions:

• Custom curated dataset of potato leaf diseases from Bangladeshi farms
• Transfer learning with fine-tuned CNN architectures optimized for mobile deployment
• Farmer-facing guidance module providing actionable disease-specific treatment recommendations

Status: Government-funded research — manuscript in preparation

Motivation: Predictive models for clinical risk assessment often sacrifice interpretability for accuracy. In healthcare, understanding why a model predicts high risk is as important as the prediction itself.

Approach: We integrate physics-informed constraints into neural network architectures for tabular medical data, producing models that respect known clinical relationships while providing feature-level explanations for their predictions.

Key contributions:

• Physics-informed regularization adapted for multi-condition medical tabular data
• Interpretable feature attribution aligned with established clinical knowledge
• Robustness evaluation across diverse patient populations and data quality levels

Status: Manuscript in preparation

Motivation: Gestational diabetes is usually diagnosed at 24–28 weeks, after much of the chance for early prevention has passed. Many machine-learning studies report near-perfect accuracy, but most use small single-hospital datasets, include late-pregnancy tests such as the glucose tolerance test, and are rarely validated on new data, so their real early-prediction value is unclear.

Approach: Using 6.9 million US births (CDC NVSS natality data), we predict gestational diabetes from 24 variables available at or before the first trimester, train on 2023 and test once on the unseen 2024 cohort. An FT-Transformer is benchmarked against logistic regression, XGBoost, LightGBM, CatBoost, MLP, TabNet and TabPFN, with calibration, decision-curve and fairness analyses. Predictions are explained with SHAP at four levels — global, individual patient, exact pairwise interactions and population subgroups — and the explanations are cross-checked against a second model and against attention weights. We also audit a widely used public GDM dataset (3,525 records) with repeated nested cross-validation.

Key findings:

• Early prediction from routine first-trimester data reaches AUC 0.71 on the future-year test set; the FT-Transformer ties with gradient-boosted trees, so its value lies in transparent, calibrated risk estimates rather than higher accuracy
• Maternal age and pre-pregnancy BMI drive most predictions, and the model relies on the same risk factors in every racial/ethnic and age group; feature interactions are small
• SHAP explanations are stable and model-agnostic (rank agreement 0.96 with XGBoost), whereas transformer attention weights agree only partially and are not a reliable explanation
• In the public benchmark dataset, the pattern of missing values alone predicts the diagnosis (AUC 0.98), showing that the near-perfect accuracies reported on it largely reflect how the data were collected

Status: Manuscript in preparation for journal submission

Motivation: Accurate brain tumor segmentation from multi-modal MRI is critical for surgical planning, yet existing models lack uncertainty estimates that clinicians need to assess prediction reliability.

Approach: We adapt foundation models for volumetric medical image segmentation, incorporating voxel-level uncertainty quantification to highlight regions where the model is less confident — enabling clinicians to identify areas requiring manual review.

Key contributions:

• Foundation model adaptation strategy for 3D medical image segmentation
• Voxel-level uncertainty maps for clinical decision support in surgical planning
• Interpretable attention visualization aligned with radiological features

Status: Manuscript in preparation