Federated Sepsis Prediction With Privacy Analysis

ISEF Category: Computational Biology and Bioinformatics

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A model can look smart on one hospital’s data and fail at the next one. That gap matters when the stakes are sepsis, where every hour can change outcomes. Federated learning tries to fix that by training across many sites without pooling raw patient records. Your job is to test how well that tradeoff works, and how much privacy costs in accuracy.

What Is It?

Federated learning is a way to train one machine-learning model across many data holders without moving all the data into one place. Think of it like group studying. Each hospital trains on its own notes, then sends back model updates instead of patient records. A central server combines those updates into a shared model.

This project turns real hospital data into a set of "virtual hospitals." You split a public ICU database into separate groups that mimic different sites, then train and compare models under different data-sharing rules. Sepsis early warning means predicting which patients may develop sepsis soon, so the model can flag risk before things get worse. Differential privacy adds noise to protect individual records. That noise helps privacy, but it can also weaken prediction quality.

Why This Is a Good Topic

This is a strong science fair topic because you can test a real research question with public data, clear metrics, and a meaningful tradeoff. You can measure how model accuracy changes when data gets split across sites, when the splits become less even, and when privacy noise increases. That gives you a direct link between AI, privacy, and healthcare. You also get to learn model evaluation, class imbalance, ROC curves, and statistical comparison, which are useful skills for serious research.

Research Questions

• How does splitting MIMIC-IV data into more virtual hospitals change sepsis prediction performance? ?
• What is the effect of non-identical hospital data distributions on federated model accuracy? ?
• Does adding differential privacy noise reduce sepsis early-warning sensitivity more than specificity? ?
• To what extent does a federated model outperform a local-only model on held-out virtual hospital data? ?
• Which privacy budget gives the best balance between utility and privacy loss? ?
• How does the timing window for sepsis prediction affect AUC, sensitivity, and false alarm rate? ?

Basic Materials

• A computer with Python support and enough memory for tabular machine-learning work.
• Access to the public MIMIC-IV database through PhysioNet and the required data use agreement.
• A spreadsheet or note-taking app for tracking experiments and versioned data splits.
• Python libraries such as pandas, scikit-learn, numpy, and matplotlib.
• A simple plotting tool for confusion matrices, ROC curves, and calibration charts.
• A code editor such as VS Code or Jupyter Notebook.
• A secure folder system for storing data dictionaries, scripts, and experiment logs.

Advanced Materials

• A workstation with a strong CPU, 32 GB or more of RAM, and optional GPU support for faster training.
• Python federated-learning libraries such as Flower, FedML, or PySyft.
• A reproducible computing setup such as Docker or conda environments.
• Access to MIMIC-IV concepts, itemized event tables, and derived cohort scripts.
• Differential privacy tools or libraries for adding and tracking privacy noise.
• Statistical analysis software for permutation tests, bootstrap confidence intervals, and calibration analysis.
• A cloud or university server for running repeated training trials.

Software & Tools

• Python: Handles data cleaning, model training, and metric calculations for tabular ICU data.
• Jupyter Notebook: Lets you document code, figures, and experiment decisions in one place.
• scikit-learn: Builds baseline classifiers and computes standard evaluation metrics.
• Flower: Simulates federated training across multiple virtual hospitals.
• Matplotlib: Plots ROC curves, confusion matrices, and performance changes across privacy budgets.

Experiment Steps

1. Define the prediction task, including the outcome window, the input window, and the patient cohort you will study.
2. Partition the public ICU data into virtual hospitals that differ in size, case mix, or patient characteristics.
3. Build a strong local-only baseline so you can compare federated learning against a simple reference point.
4. Train a federated model and test how performance changes when the sites are balanced versus uneven.
5. Add differential privacy settings and compare the privacy budget against key metrics such as sensitivity, specificity, and AUC.
6. Check whether your model stays well calibrated and whether any site group loses performance more than the others.

Common Pitfalls

• Using random row splits instead of patient-level splits, which leaks information between train and test sets.
• Ignoring class imbalance, which can make a weak sepsis predictor look strong because most cases are negative.
• Mixing feature windows that happen after the prediction time, which creates data leakage and fake performance.
• Treating every virtual hospital as identical, which hides the effect of non-IID data on federated learning.
• Changing too many model settings at once, which makes it hard to tell whether privacy noise or data partitioning caused the result.

What Makes This Competitive

A class-level version of this project stops at a basic federated model and a single accuracy score. A stronger entry compares multiple partition schemes, tests more than one privacy budget, and reports uncertainty with confidence intervals. You can also separate overall performance from subgroup performance, which matters when one virtual hospital behaves very differently from the others. Careful calibration checks and false alarm analysis can push the project toward real research quality.

Project Variations

• Use a different ICU task, such as predicting acute kidney injury or in-hospital mortality, and compare the privacy tradeoff.
• Change the partition rule so virtual hospitals differ by age, sex, or admission source, then test fairness across sites.
• Compare federated learning with centralized training, local-only training, and transfer learning on the same public dataset.

Learn More

• PhysioNet and MIMIC-IV documentation: Find the public ICU database, data dictionaries, and access steps on PhysioNet.
• NIH PubMed: Search for review articles on federated learning in healthcare, differential privacy, and sepsis prediction.
• Nature Medicine and npj Digital Medicine: Search these journals for recent peer-reviewed papers on privacy-preserving medical AI.
• MIT OpenCourseWare: Look for free machine-learning and statistics course materials that cover model evaluation and regularization.
• scikit-learn User Guide: Read the free documentation for classification metrics, calibration, and cross-validation.
• Flower Documentation: Use the free federated-learning docs to learn how simulated clients and server rounds work.