Predicting ICU Kidney Injury

ISEF Category: Biomedical and Health Sciences

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

The Hook

A kidney can start failing before anyone writes the diagnosis down. In the ICU, the warning signs hide in hours of vitals and lab results. A temporal transformer reads that timeline like a movie, not a snapshot. Your project can test how early the pattern appears and which signals push the model to flag risk.

What Is It?

Acute kidney injury, or AKI, happens when the kidneys suddenly lose filtering power. In ICU data, the warning signs do not show up in one clean measurement. They unfold across a timeline of blood pressure, heart rate, urine output, and lab values, like clues in a long case file.

A temporal transformer is a machine learning model that reads the order of those clues. SHAP, short for Shapley Additive Explanations, shows which inputs pushed a prediction higher or lower. Together, they let you build a model that predicts risk and explains the decision in plain language.

Why This Is a Good Topic

This is a strong science fair topic because you can use public ICU data to tackle a real medical problem. You can compare a simple baseline with a temporal transformer, measure how far ahead each model catches risk, and check whether the explanations stay stable across patients. The project teaches time-series modeling, model evaluation, and clinical interpretation without needing a wet lab.

Research Questions

• How does adding lab trends change AKI prediction compared with vitals alone?
• What is the effect of using a temporal transformer instead of logistic regression on 12-hour early warning accuracy?
• Does a model trained on one ICU stay length generalize better than a model trained on a different stay length?
• To what extent do SHAP explanations stay consistent across patients with different AKI risk levels?
• Which prediction window gives the best tradeoff between early warning and false alarms?
• How does class balancing change recall for rare AKI cases?

Basic Materials

• Laptop with at least 16 GB of RAM.
• Free Python installation.
• JupyterLab or VS Code.
• PhysioNet account with MIMIC-IV access approval.
• MIMIC-IV ICU data files.
• Spreadsheet software for quick checks.
• Git for version control.

Advanced Materials

• University workstation or cloud VM with a GPU.
• Secure data environment approved for MIMIC-IV.
• Python with PyTorch, pandas, scikit-learn, and SHAP.
• JupyterHub or VS Code Remote.
• High-speed storage for cached time-series files.
• R or Python SciPy for statistical tests.
• Plotting tools for patient timelines.

Software & Tools

• Python: Cleans ICU tables, trains baselines, and runs the transformer.
• JupyterLab: Lets you inspect patient timelines and compare model runs.
• PyTorch: Trains the temporal transformer on sequential vitals and labs.
• SHAP: Turns model outputs into feature-attribution plots clinicians can read.
• pandas: Reshapes raw ICU tables into per-patient time series.

Experiment Steps

1. Define the AKI label and prediction window you will use.
2. Choose one patient cohort and one exclusion rule set so the data stays comparable.
3. Decide how you will turn irregular vitals and labs into fixed-length time windows.
4. Build a simple baseline first, then compare it with the temporal transformer.
5. Plan evaluation metrics that reward early warning, not just overall accuracy.
6. Map SHAP outputs to a short clinician-facing summary for each prediction.

Common Pitfalls

• Mixing measurements taken after the AKI label, which leaks the answer into training.
• Keeping patients with almost no labs, which makes the model learn who gets tested instead of who gets sick.
• Using overall accuracy on a rare outcome, which makes a weak model look better than it is.
• Comparing predictions from different time windows without matching the same lead time, which inflates early-warning claims.
• Treating a noisy SHAP plot as a medical explanation, which hides unstable feature attributions across runs.

What Makes This Competitive

A stronger entry would compare more than one model and more than one time horizon, then show where the gains really come from. You can raise the bar by testing calibration, not just AUC, and by checking whether the explanation patterns stay similar across subgroups. A clean, clinician-readable explanation layer makes the work feel like a real decision-support study, not just a coding demo. If you add careful error analysis for missed AKI cases, the project becomes much harder to dismiss.

Project Variations

• Restrict the cohort to sepsis patients and test whether AKI warning signs change in that subgroup.
• Compare a temporal transformer with XGBoost, LSTM, and logistic regression on the same patient windows.
• Use lab-only, vitals-only, and combined inputs to measure which signal gives the earliest warning.

Learn More

• PhysioNet MIMIC-IV: Find the ICU database documentation, access steps, and data dictionary on PhysioNet.
• NIH NIDDK Acute Kidney Injury: Read the plain-language overview and clinical background on the NIDDK site.
• KDIGO Clinical Practice Guideline for AKI: Search the guideline PDF and summary articles for the standard AKI definition.
• SHAP Documentation: Read how SHAP values explain model predictions and how to make summary plots.
• PubMed: Search review articles on acute kidney injury prediction, ICU time-series models, and model interpretability.