In a test cohort of 11,523 hospital encounters, only 242 blood-culture episodes were pathogen-positive. A machine-learning model placed 81 of those positives inside the 577 encounters with the highest predicted risk. In other words, the top 5% contained one-third of all positives—but 496 of the 577 flagged encounters were still negative.[1]
That combination is the right way to read a low-prevalence medical model. The system concentrated risk meaningfully: positivity rose from 2.10% across the test cohort to 14.0% in the top group, a 6.68-fold enrichment. It did not turn a blood-culture question into a diagnosis.[1]
The high-risk group was useful and still mostly wrong
Primary test result
| Measure | Value | Context |
|---|---|---|
| Test encounters | 11,523 | The evaluation cohort contained 242 pathogen-positive episodes.Source 1 |
| Observed prevalence | 2.10% | This low base rate makes precision difficult even with useful discrimination.Source 1 |
| AUROC | 0.819 | The 95% confidence interval was 0.792–0.844.Source 1 |
| Average precision | 0.124 | This metric is more sensitive to rare-positive performance than AUROC.Source 1 |
| Top risk group | 577 encounters | The top 5% contained 81 positives and had 14.0% positivity.Source 1 |
| Mean predicted risk | 3.78% | Observed positivity was 2.10%, showing overprediction in the test cohort.Source 1 |
What the top-five-per-cent threshold produced
| Group | Encounters | Positive cultures | Observed positivity |
|---|---|---|---|
| Highest predicted 5% | 577 | 81 | 14.0%Source 1 |
| Remaining 95% | 10,946 | 161 | 1.47%Source 1 |
| All test encounters | 11,523 | 242 | 2.10%Source 1 |
The top group captured 33.5% of all positives: 81 divided by 242. It also contained 86.0% non-positive episodes: 496 divided by 577. Those are not contradictory results. One describes concentration; the other describes the burden of acting on a rare event.[1]
The model ranked episodes; it did not identify an organism
The target was whether the episode eventually yielded a pathogen-positive blood culture. The model did not name the organism, establish contamination, determine antibiotic susceptibility or prove bloodstream infection at the two-hour mark. It used electronic health-record information available after the first culture record to estimate risk.[1]
Laboratory values contributed most, particularly lymphocyte percentage. Secondary analyses considered previous culture results. Because the data came from MIMIC-IV, a de-identified record from one health system, the reported relationships reflect that institution's testing, coding, population and laboratory workflows.[1][2]
Two hours won because waiting added little
Why the earliest tested prediction point was selected
| Time point | Potential advantage | Study finding |
|---|---|---|
| 2 hours | Earlier opportunity to review testing and care | Chosen as the primary point; discrimination was already useful.Source 1 |
| 4 hours | More laboratory and clinical data | The additional gain was small and uncertain.Source 1 |
| 6 hours | Still more data | Later improvement remained too limited to outweigh delay for the study’s purpose.Source 1 |
This is an important design choice. A later model can always accumulate more clues, but diagnostic stewardship is valuable only if the information arrives while clinicians can still change what they do. The study therefore optimised a decision time, not merely a score.[1]
Calibration is the immediate deployment warning
Mean predicted risk was 3.78%, compared with 2.10% observed. A system can rank patients well and still produce probabilities that are too high. If those numbers were used directly for thresholds, staffing or treatment, overprediction could create unnecessary alerts and interventions.[1]
Possible uses and unsafe shortcuts
| Potential use after validation | Unsafe interpretation |
|---|---|
| Prioritise review of cultures and clinical context in a small high-risk group. | Treat a high score as proof of bloodstream infection.Source 1 |
| Identify low-risk groups for a prospective stewardship study. | Cancel cultures or antibiotics solely from the retrospective score.Source 1 |
| Allocate laboratory or infectious-disease attention earlier. | Assume the model identifies the organism or susceptibility.Source 1 |
| Monitor calibration as patient mix and testing policy change. | Deploy the MIMIC-IV probability without local validation and recalibration.Source 1Source 2 |
The paper's strongest result is therefore not the AUROC. It is the concentration achieved early enough to matter, accompanied by enough calibration error to prevent premature clinical use. The model could become a useful queue. It is not yet a verdict—and in a population where 97.9% of episodes are negative, the difference between those two roles is the entire safety case.[1]
Sources & further reading
- Predicting pathogen-positive blood culture episodes at two hours using electronic health record data Scientific Reports · 15 Oct 2026 · Accessed 16 Oct 2026
- MIMIC-IV clinical database PhysioNet · 31 Jan 2024 · Accessed 16 Oct 2026