MediScan Health built a binary classifier to flag patients for follow-up diabetes screening based on routine lab results and visit history. The model is now in production across 120 clinics, but clinical leaders disagree on whether the current threshold is too conservative.
| Metric | Current Model | Lower Threshold Option |
|---|---|---|
| Precision | 0.91 | 0.74 |
| Recall | 0.58 | 0.81 |
| F1 Score | 0.71 | 0.77 |
| Accuracy | 0.95 | 0.92 |
| Patients flagged per 10,000 | 640 | 1,120 |
| True cases detected per 10,000 | 582 | 812 |
| Missed positive cases per 10,000 | 418 | 188 |
The current model produces very few false alarms, but it misses a meaningful share of true diabetes-risk patients. Operations can support only a limited number of follow-up screenings each week, while clinicians argue that missed cases are more harmful than unnecessary outreach.