Sign inJoin the network

Glossary · 4 min read

Positive and negative predictive value (PPV and NPV)

What PPV and NPV mean for a clinical AI alert, why the same model can be right most of the time in one unit and rarely in another, and the arithmetic to check a vendor's numbers yourself. As of September 2026.

By AIMOCSUpdated

The short version

  • PPV is the share of positive results that are truly positive; NPV is the share of negative results that are truly negative. For an AI alert, PPV answers: when this fires, how often is it right?
  • Both depend on prevalence. An illustrative alert with 80% sensitivity and 90% specificity has a PPV of 14% at 2% prevalence and 67% at 20% prevalence.
  • A widely deployed sepsis model flagged 6,971 of 38,455 hospitalizations and caught 843 of 2,552 sepsis cases, a PPV of about 12%, or roughly eight alerts per true case.
  • When a condition is rare, a high NPV is close to automatic: a sepsis model that never fired would have scored an NPV of 93% in the same cohort.
On this page

Positive predictive value (PPV) is the share of patients with a positive result who truly have the condition. Negative predictive value (NPV) is the share with a negative result who truly do not 1. For a clinical AI alert, PPV answers the question a clinician asks at the bedside: when this fires, how often is it right?

Why do PPV and NPV matter in healthcare?

Sensitivity and specificity describe a test. PPV and NPV describe what a result means for the patient in front of you, and they depend on how common the condition is in the population being tested 1. The same model can have a useful PPV in an ICU and a poor one on a general ward.

For alerting systems, PPV also sets the workload. Its reciprocal is the number needed to evaluate: how many patients a team must work up to find one true case. A PPV of 20% means five evaluations per true positive 2.

How does prevalence change PPV?

Take an illustrative sepsis alert with 80% sensitivity and 90% specificity, applied to 1,000 patients.

General ward, 2% prevalence. 20 patients have sepsis and the alert catches 16. Of the 980 without sepsis, 10% (98) trigger false alarms. PPV = 16 / (16 + 98) = 14%. NPV = 882 / (882 + 4) = 99.5%.

ICU cohort, 20% prevalence. 200 patients have sepsis and the alert catches 160. Of the 800 without, 80 trigger false alarms. PPV = 160 / (160 + 80) = 67%. NPV = 720 / (720 + 40) = 94.7%.

Same model, same threshold. PPV rose from 14% to 67% because prevalence rose, and NPV moved the other way.

What does this look like in a real deployment?

An external validation of a widely implemented proprietary sepsis model at one academic health system gives the counts to check. Sepsis occurred in 2,552 of 38,455 hospitalizations (7%). At its alert threshold, the model flagged 6,971 hospitalizations and missed 1,709 sepsis cases 3.

  • True positives: 2,552 − 1,709 = 843, a sensitivity of 33%.
  • PPV: 843 / 6,971 = 12%, or about eight alerts for every true case.
  • NPV: of 31,484 unflagged hospitalizations, 1,709 had sepsis, so 29,775 / 31,484 = 95%.

That NPV sounds reassuring until you notice that 35,903 of the 38,455 hospitalizations (93%) had no sepsis. A model that never fired at all would have scored an NPV of 93%.

Common misunderstandings

PPV is a fixed property of the model. It moves with prevalence 1. A PPV measured in a high-prevalence validation cohort will overstate PPV in a lower-prevalence population.

Sensitivity and specificity, at least, stay put. Less than often assumed. Across 23 meta-analyses, specificity tended to be lower where prevalence was higher, likely because patient spectrum differs between settings 4.

A low PPV means the tool is useless. For early warning scores, where a positive triggers a brief assessment, a PPV below 50% can be acceptable 2. The real question is whether the team can absorb the workload without tuning alerts out.

A high NPV proves the model is safe. When a condition is rare, NPV is high almost by default, as the sepsis example shows.

What to ask a vendor

  • What was the prevalence in your validation cohort, and how does it compare with ours?
  • What are PPV and NPV at the exact threshold we will use? An AUROC alone cannot answer this.
  • How many alerts per 100 patients per day should we expect? Romero-Brufau and colleagues recommend reporting this rate alongside sensitivity 2.
  • Were these numbers measured at an external site?

Questions and answers

  • What is the difference between PPV and sensitivity?

    Sensitivity starts from patients who have the condition and asks what share the test catches. PPV starts from patients who test positive and asks what share truly have the condition. Sensitivity describes the test; PPV describes what a positive result means in a particular population, so it changes with prevalence.

  • Why is the PPV of an AI alert often low?

    Because most alerts screen for uncommon events. When only a few percent of patients have the condition, even a fairly specific model produces more false alarms than true ones. In one external validation of a widely used sepsis model, about 12% of alerted hospitalizations were true sepsis cases.

  • What is a good PPV for a clinical alert?

    It depends on the stakes and on what an alert triggers. For early warning of deterioration, where a positive leads to a brief bedside assessment, methodologists argue a PPV well below 50% can be acceptable. The useful framing is workload: a PPV of 20% means five evaluations for every true case.

Sources

  1. Altman DG, Bland JM. Diagnostic tests 2: Predictive values. BMJ. 1994;309(6947):102. doi.org/10.1136/bmj.309.6947.102
  2. Romero-Brufau S, Huddleston JM, Escobar GJ, Liebow M. Why the C-statistic is not informative to evaluate early warning scores and what metrics to use. Critical Care. 2015;19:285. doi.org/10.1186/s13054-015-0999-1
  3. Wong A, Otles E, Donnelly JP, et al. External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients. JAMA Internal Medicine. 2021;181(8):1065-1070. doi.org/10.1001/jamainternmed.2021.2626
  4. Leeflang MMG, Rutjes AWS, Reitsma JB, Hooft L, Bossuyt PMM. Variation of a test's sensitivity and specificity with disease prevalence. CMAJ. 2013;185(11):E537-E544. doi.org/10.1503/cmaj.121286

Every claim on this page is tied to a numbered primary source above. Read how we source and review at our editorial policy.