Transparent Clinical Risk Tool With SHAP Explanations

ISEF Category: Translational Medical Science

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

The Hook

Doctors make life-or-death decisions fast, but most AI tools act like black boxes. Your project asks a simple question, can a risk model explain itself in a way clinicians can actually use? That mix of prediction, explanation, and usability makes this a strong science fair idea. It also connects directly to safer care.

What Is It?

This project builds a clinical decision-support prototype. That means a program that takes in patient data and gives a risk estimate, like a warning light on a dashboard. The twist is transparency. Instead of just saying, “high risk,” the tool also shows which features pushed the prediction up or down.

SHAP stands for Shapley Additive Explanations. In plain language, SHAP tries to split a model’s prediction into pieces, so you can see how much each input mattered. If a patient case gets a high-risk score, SHAP can show whether age, blood pressure, lab values, or other features drove that result. You are not building a diagnosis machine. You are testing whether an explainable interface makes model output easier to trust, understand, and use.

Why This Is a Good Topic

This is a strong topic because you can test both the technical model and the human response to it. The data analysis is real, but the project still has a clear end point, a tool that makes a prediction and explains it. That gives you a concrete product and measurable outcomes. It also connects to a real problem in medicine, which is that clinicians need fast, transparent tools, not mysterious scores.

Research Questions

• How does adding SHAP explanations change clinician usability scores for the same risk prediction model?
• What is the effect of different explanation layouts on the speed of clinical interpretation?
• Does a transparent model increase user trust compared with a black-box version with the same accuracy?
• To what extent do feature explanations help users identify the main risk drivers in a case vignette?
• Which case presentation format produces the lowest disagreement between the model and mentor-clinician ratings?
• How does model performance change when you test it on case sets from different patient subgroups?

Basic Materials

• Laptop or desktop computer with a modern browser.
• Python installed with a code editor such as VS Code or Jupyter.
• Public, de-identified clinical dataset access plan, such as MIMIC-IV after training and data use approval.
• Spreadsheet software for tracking cases, labels, and reviewer scores.
• Survey form tool for clinician usability ratings.
• Documented case-vignette template for human testing.
• GitHub account for version control and sharing code.

Advanced Materials

• University or mentor-guided access to MIMIC-IV or a similar de-identified electronic health record dataset.
• Python environment with scikit-learn, XGBoost, SHAP, pandas, NumPy, and matplotlib.
• Secure compute environment approved for protected health data workflows.
• REDCap or another research survey platform for clinician usability scoring.
• Statistics software for agreement testing, calibration analysis, and subgroup checks.
• Optional web framework such as Streamlit or Flask for the browser-based prototype.
• IRB or mentor review documents, if human subject review is required.

Software & Tools

• Python: Runs the modeling, data cleaning, and explanation code for the prototype.
• SHAP: Calculates feature contributions so you can show why the model made each prediction.
• Streamlit: Builds a simple browser interface for entering cases and viewing results.
• pandas: Organizes case data and helps you prepare features for analysis.
• scikit-learn: Trains baseline models and supports validation, calibration, and metrics.

Experiment Steps

1. Define the decision task you want the tool to support, such as risk ranking, triage, or readmission prediction.
2. Choose one clinical dataset format and one target outcome so your project stays focused.
3. Build a baseline model first, then add an explanation layer so you can compare both versions fairly.
4. Design a user interface that shows the prediction, the top features, and a short summary of model confidence.
5. Plan a clinician or mentor review process with consistent scoring rubrics for usability, trust, and clarity.
6. Decide in advance how you will test subgroup performance, explanation quality, and agreement between reviewers.

Common Pitfalls

• Using messy labels from the dataset without checking how the outcome was defined, which can make the model learn the wrong target.
• Testing the tool only on one patient type, which hides poor performance in other groups.
• Showing too many SHAP features at once, which makes the explanation harder to read than the raw score.
• Mixing training data and evaluation cases, which inflates accuracy and gives a false sense of quality.
• Asking reviewers vague questions like “Do you like it?”, which produces unusable usability data.

What Makes This Competitive

A competitive version needs more than a working demo. You need a clean evaluation plan, strong validation, and a clear reason your interface helps. The best projects compare explanation styles, test calibration, and check whether the tool behaves differently across patient subgroups. If you add a thoughtful user study with clinician feedback, your project starts to look like applied research, not just software.

Project Variations

• Build a simpler version that predicts one outcome, such as ICU transfer risk, instead of a broad clinical score.
• Compare SHAP with another explanation method, such as LIME, to see which one users understand better.
• Test the same prototype on different case types, such as emergency room, inpatient, or outpatient examples, to study transferability.

Learn More

• MIMIC-IV documentation: Search the PhysioNet site for the MIMIC-IV dataset, data dictionary, and access instructions.
• PhysioNet: Search for open critical care datasets and related methods papers.
• NIH PubMed: Search review articles on explainable AI in medicine, clinical decision support, and SHAP.
• SHAP documentation: Read the official documentation for explanation methods, examples, and plotting tools.
• MIT OpenCourseWare: Search for machine learning, data science, and health informatics course materials.
• Nature Medicine and npj Digital Medicine: Search for peer-reviewed papers on clinical AI evaluation and explainability.