ReadmitIQ

Predict who comes back before they leave — ML readmission risk at the point of discharge.

Comment

Here's a full Devpost project story ready to paste in:

Inspiration

Hospital readmissions are one of the most visible and costly failure points in American healthcare. Nearly 1 in 5 Medicare patients is readmitted within 30 days of discharge, costing the system an estimated $26 billion annually in avoidable spending. More importantly, many of these readmissions represent genuine gaps in care — a patient discharged without proper follow-up, a medication reconciliation that never happened, a social determinant of health that nobody flagged.

At UC Davis Medical Center, discharge nurses manage dozens of patients simultaneously with no systematic way to prioritize who needs the most attention before leaving. We built ReadmitIQ to change that — putting a real-time, explainable risk score in the hands of the care team at the exact moment it matters most: discharge planning.

What it does

ReadmitIQ is a discharge planning dashboard that predicts each patient's probability of 30-day readmission using structured clinical data available at the time of discharge. For every patient, it surfaces:

  • A risk score (0–100%) derived from an XGBoost classifier trained on MIMIC-IV
  • The top 3 clinical drivers of that score, powered by SHAP (SHapley Additive exPlanations), so clinicians understand why the model flagged a patient — not just that it did
  • An intervention checklist that automatically suggests follow-up actions: social work consults, pharmacy reconciliation, PCP callbacks, and home health referrals

The model uses 12 features extractable from any standard EHR at discharge:

$$\text{Risk} = f(\text{LOS},\ n_{\text{dx}},\ n_{\text{prev_admits}},\ \text{creatinine},\ \text{Hgb},\ \text{WBC},\ \text{age},\ \text{disposition},\ \ldots)$$

Risk tiers are defined as:

$$\text{Tier} = \begin{cases} \text{High} & p \geq 0.70 \ \text{Medium} & 0.40 \leq p < 0.70 \ \text{Low} & p < 0.40 \end{cases}$$

How we built it

Data: We used the MIMIC-IV clinical dataset (PhysioNet), joining the admissions, patients, diagnoses_icd, labevents, and services tables to construct a patient-level feature matrix with a binary 30-day readmission label.

Model: We trained an XGBoost classifier with 5-fold stratified cross-validation, using scale_pos_weight to handle class imbalance (approximately 15% readmission rate). We tuned the decision threshold to prioritize recall for high-risk patients, targeting $\text{Recall} \geq 0.70$ at the cost of some precision — the right clinical tradeoff when a missed high-risk patient is more costly than a false alarm.

Explainability: We integrated the SHAP library to generate per-patient feature attributions, surfacing the top 3 drivers for every prediction. This was non-negotiable for clinical trust — a black-box score is useless to a discharge nurse.

API: A Flask backend serves the model via a POST /predict endpoint, accepting a patient JSON and returning the risk score plus SHAP values in under 200ms.

Frontend: A React dashboard presents a sortable patient list, a semicircular risk gauge, a SHAP waterfall chart, and an intervention checklist — all designed around the discharge nurse's workflow.

Challenges we ran into

Clinical validity vs. model performance. Early feature sets with high AUC included variables not reliably available at discharge time (post-discharge lab values, follow-up appointment status). Our medical students were essential in auditing every feature for real-world availability and clinical plausibility — a step that meaningfully changed our feature set.

Class imbalance. With only ~15% of admissions resulting in readmission, naive models learned to predict "no readmission" for everyone. We addressed this through stratified cross-validation, scale_pos_weight tuning, and careful threshold selection guided by a precision-recall curve rather than accuracy.

Making SHAP interpretable to clinicians. Raw SHAP values are unintuitive. We worked with our clinical team to translate them into natural-language driver statements ("3 prior admissions in 90 days increases risk significantly") rather than displaying raw numbers.

Time. 24 hours is not a lot of time to train, evaluate, build, and polish. We made deliberate scope cuts — no authentication layer, no live EHR connection — to ship a compelling demo over a technically complete but undemonstrable system.

Accomplishments that we're proud of

  • Achieved AUC-ROC of 0.74 on the MIMIC-IV held-out test set, competitive with published readmission models that took months to develop
  • Built a fully functional end-to-end pipeline — raw EHR data to interactive clinical dashboard — in under 20 hours
  • Designed an intervention checklist grounded in evidence-based discharge protocols, reviewed and approved by our medical student team members
  • Made the model genuinely explainable: every prediction comes with a plain-English reason, not just a number

What we learned

The most important thing we learned had nothing to do with machine learning: clinical workflow is the hardest design problem. A risk score that interrupts a nurse's existing process will be ignored. A risk score embedded in the discharge workflow, at the right moment, with actionable next steps, might actually change behavior.

We also learned that medical students and engineers think about data completely differently. Engineers optimized for AUC. Clinicians asked "but what does this mean for Mrs. Johnson?" That tension made the project significantly better.

On the technical side: SHAP explainability is not optional for clinical ML. It is the feature.

What's next for ReadmitIQ

  • EHR integration via FHIR R4 — connecting directly to Epic or Cerner so risk scores populate automatically without manual data entry
  • Prospective validation — running the model in shadow mode alongside standard discharge planning at UC Davis Health to measure real-world calibration
  • Expanded feature set — incorporating social determinants of health (housing instability, food insecurity scores) from structured SDOH screening data
  • Feedback loop — capturing whether flagged interventions were completed and whether readmission occurred, enabling continuous model retraining
  • CMS compliance framing — mapping high-risk flags to the Hospital Readmissions Reduction Program (HRRP) penalty conditions (CHF, COPD, pneumonia, hip/knee replacement) to prioritize institutional ROI

Built With

  • ai
  • claude
Share this project:

Updates