A hospital model does not need to diagnose every bloodstream infection to change a workflow. It can be useful if it concentrates a scarce outcome into a smaller group that clinicians can review more closely. In a held-out set of 11,523 encounters, a new model placed 81 of the 242 positive blood cultures inside its highest-risk 5%.[1]
That top band contained 577 encounters, so its positivity rate was 14.0%. The overall held-out rate was 2.10%, making the band about 6.68 times more concentrated than the test population. The same numbers also show the limit: 496 of the 577 highest-risk encounters still had no positive culture.[1]
Held-out performance
| Measure | Value | Context |
|---|---|---|
| Test encounters | 11,523 | The held-out set contained 242 positive cultures.Source 1 |
| AUROC | 0.819 | 95% confidence interval 0.792–0.844; measures ranking across thresholds.Source 1 |
| Average precision | 0.124 | More informative than accuracy when the positive class is rare.Source 1 |
| Top 5% band | 81 positives among 577 encounters | A 14.0% positivity rate and 6.68-fold enrichment over the test prevalence.Source 1 |
| Calibration | 3.78% predicted vs 2.10% observed | The model overestimated absolute risk in the held-out period.Source 1 |
Ranking and probability are different jobs
AUROC asks whether a randomly selected positive encounter tends to receive a higher score than a randomly selected negative one. Calibration asks whether a prediction such as 10% actually corresponds to about one positive result in ten similar cases. This model ranked reasonably well but predicted a higher mean risk than the observed test rate.[1]
That distinction changes how the score should be displayed. A poorly calibrated “18% chance” can mislead a clinician even when the same model is useful for ordering a work queue. Local recalibration may preserve ranking while correcting the probabilities for a new hospital or time period.
The model learned from routine information
The dataset contained 145,914 encounters, including 6,404 positive cultures. Laboratory values were important features, with lymphocyte percentage among the strongest contributors. Models evaluated at later timepoints gained only modestly from additional information, suggesting much of the usable signal was already available around the time the culture was ordered.[1]
Three possible uses—and their risks
| Use | Potential value | Failure to guard against |
|---|---|---|
| Prioritise review | Bring the highest-risk encounters to attention sooner. | Treating the score as proof of infection.Source 1 |
| Antimicrobial-stewardship support | Identify cases where additional scrutiny is warranted. | Delaying empiric treatment in a clinically unstable patient.Source 1 |
| Laboratory workflow planning | Anticipate a higher-yield subset of specimens. | Changing collection quality or culture decisions solely from the model.Source 1 |
Prospective evidence is the missing step
The study is retrospective. A real deployment must show that the score reaches the right person at the right time, changes a decision safely and remains calibrated as pathogens, ordering practice and patient mix change. The headline result is promising because it compresses risk. It is not permission to skip the culture that provides the diagnosis.
Sources & further reading
- Machine learning prediction of blood culture positivity at the time of ordering Scientific Reports · 15 Oct 2026 · Accessed 16 Oct 2026