MediScan, a digital health company, uses a binary classification model to flag patients for follow-up diabetes screening after an annual checkup. The current debate is whether to deploy a model version with higher precision or one with higher recall, because missed high-risk patients can delay treatment, while too many false alarms overwhelm clinicians.
Two candidate logistic regression models were evaluated on the same validation set of 20,000 patients, with 1,000 actual positive cases.
| Metric | Model A | Model B |
|---|---|---|
| Precision | 0.91 | 0.62 |
| Recall | 0.48 | 0.84 |
| F1 Score | 0.63 | 0.71 |
| Accuracy | 0.96 | 0.94 |
| AUC-ROC | 0.88 | 0.89 |
| Flagged patients | 527 | 1,355 |
The clinical operations team can handle at most 1,400 follow-up screenings per month. Each unnecessary follow-up costs about $40 in staff time and patient outreach. Each missed high-risk patient is estimated to create $900 in downstream medical and retention costs.