biomed-agent-kit

Process and analyze biomedical text data (clinical notes, radiology reports, PubMed abstracts). Use when: (1) Loading biomedical language models (PubMedBERT, ClinicalBERT, BioBERT, BioGPT) and extracting embeddings, (2) Named entity recognition for medical concepts with scispaCy, (3) UMLS concept linking, (4) Text classification of clinical documents, (5) Extracting structured data from unstructured clinical text (vitals, lab values, findings).

Load and preprocess biomedical data (DICOM, NIfTI, whole-slide images, tabular clinical data). Use when: (1) Loading DICOM/NIfTI/WSI files and extracting metadata, (2) Intensity normalization for CT/MRI, (3) Stain normalization for histopathology, (4) Building PyTorch Dataset classes with caching or HDF5, (5) Designing augmentation pipelines for medical imaging (2D or 3D).

Deploy and optimize trained biomedical ML models for inference. Use when: (1) Exporting models to ONNX or TorchScript, (2) Building inference pipelines with sliding window for 3D volumes or patch-based prediction for WSI, (3) Test-time augmentation (TTA), (4) Optimizing inference speed with torch.compile or quantization, (5) Packaging models as REST APIs or Docker containers for serving.

Evaluate biomedical ML models with appropriate metrics and confidence intervals. Use when: (1) Computing classification metrics (AUC-ROC, balanced accuracy, sensitivity, specificity, F1) with confidence intervals, (2) Evaluating segmentation models (Dice, IoU, Hausdorff, surface Dice), (3) Survival analysis (C-index, Kaplan-Meier, Cox PH, time-dependent AUC), (4) Statistical comparison between models (Wilcoxon, paired t-test), (5) Calibration assessment (Brier score, ECE, reliability diagrams), (6) Regression metrics (MAE, RMSE, R-squared, Bland-Altman), (7) Multi-label classification metrics.

Run systematic ML experiments with production-grade patterns. Use when: (1) Setting up experiment grids with cross-validation, (2) Managing GPU memory, multi-GPU worker pools, or OOM protection, (3) Designing patient-level or site-aware data splits, (4) Tracking experiment completion with resumability, (5) Distributing work across GPUs, (6) Hyperparameter tuning with Optuna.

Research biomedical and ML papers across the internet, focusing on top venues and publishers. Use when: (1) Searching for state-of-the-art methods in a biomedical domain, (2) Finding papers from top venues (Nature Medicine, Lancet, MICCAI, NeurIPS, ICML, CVPR, TMI, MedIA), (3) Reviewing related work for a research topic, (4) Finding benchmark datasets or baselines for a task, (5) Comparing methods across papers, (6) Summarizing recent advances in a biomedical subfield.

Integrate published GitHub repositories into the current project. Use when: (1) Adding a pretrained model or method from a paper's GitHub repo, (2) Wrapping an external tool as a feature extractor or preprocessing step, (3) Resolving dependency conflicts between an external repo and the current project, (4) Adapting an external repo's data format to match the current pipeline, (5) Vendoring or submoduling code from another repository, (6) Loading models from HuggingFace Hub, MONAI Model Zoo, or timm.

Generate publication-quality figures and markdown reports for biomedical ML experiments. Use when: (1) Creating heatmaps, forest plots, radar charts, or sensitivity-specificity scatter plots, (2) Plotting ROC curves, precision-recall curves, or Kaplan-Meier survival curves, (3) Writing structured experiment reports, (4) Exporting results to LaTeX tables for paper submission, (5) Producing figures meeting publication DPI and formatting standards.

Train and fine-tune biomedical ML models with production-grade patterns. Use when: (1) Writing training loops with early stopping and LR scheduling, (2) Transfer learning and fine-tuning pretrained models (freeze/unfreeze, discriminative LR), (3) Mixed precision training (fp16/bf16), (4) Gradient accumulation for large effective batch sizes, (5) Checkpoint management (save/load best model, resume training).

Bootstrap new biomedical ML research repositories with production-grade structure. Use when: (1) Creating a new research project from scratch, (2) Setting up pyproject.toml with uv, (3) Establishing directory layout, CLAUDE.md, or config patterns for a biomedical/healthcare ML project.