Multimodal Survival Prediction Model
Budget: $30 – $250 USD
I need a robust deep-learning pipeline that predicts patient survival by fusing multiple modalities of medical records. The core dataset consists of structured EHR fields plus unstructured clinical notes and reports; no imaging or sensor streams are involved for this phase.
Your job is to design, implement, and validate a multimodal architecture that can ingest the tabular variables alongside free-text notes, learn the complementary signals in each, and output calibrated survival probabilities (e.g., 30-, 90-, 365-day). Feel free to combine techniques such as Transformer-based NLP encoders, attention-based fusion layers, and survival-specific loss functions like Cox partial likelihood or DeepSurv variants—as long as the final model is reproducible and explainable.
Key deliverables
• Clean, well-documented preprocessing scripts for both structured fields and clinical text
• The complete training pipeline (PyTorch or TensorFlow preferred) with modular code for experimentation
• Evaluation report covering discrimination (C-index, AUC) and calibration, plus ablation results showing fusion benefits over single-modality baselines
• Inference notebook or API snippet that takes new patient records and returns survival curves/probabilities
• Brief write-up (max 3 pages) detailing model choices, hyper-parameters, and clinical interpretability methods (e.g., SHAP on tabular features, attention heatmaps for text)
I will provide a de-identified dataset, feature dictionary, and annotation guidelines for the notes. Please estimate timeline and any additional data requirements you foresee so we can lock in milestones quickly.
Your job is to design, implement, and validate a multimodal architecture that can ingest the tabular variables alongside free-text notes, learn the complementary signals in each, and output calibrated survival probabilities (e.g., 30-, 90-, 365-day). Feel free to combine techniques such as Transformer-based NLP encoders, attention-based fusion layers, and survival-specific loss functions like Cox partial likelihood or DeepSurv variants—as long as the final model is reproducible and explainable.
Key deliverables
• Clean, well-documented preprocessing scripts for both structured fields and clinical text
• The complete training pipeline (PyTorch or TensorFlow preferred) with modular code for experimentation
• Evaluation report covering discrimination (C-index, AUC) and calibration, plus ablation results showing fusion benefits over single-modality baselines
• Inference notebook or API snippet that takes new patient records and returns survival curves/probabilities
• Brief write-up (max 3 pages) detailing model choices, hyper-parameters, and clinical interpretability methods (e.g., SHAP on tabular features, attention heatmaps for text)
I will provide a de-identified dataset, feature dictionary, and annotation guidelines for the notes. Please estimate timeline and any additional data requirements you foresee so we can lock in milestones quickly.