Paediatric Index of Mortality 3 (PIM3)

PIM3 score (logit) = 3.8233 × pupils − 0.5378 × elective + 0.9763 × ventilated + 0.0671 × |base excess| − 0.0431 × SBP + 0.1716 × (SBP² ÷ 1000) + 0.4214 × (FiO₂/PaO₂ term) − 1.2246 × bypass-cardiac recovery − 0.8762 × non-bypass-cardiac recovery − 1.5164 × non-cardiac recovery + 1.6225 × very-high-risk diagnosis + 1.0725 × high-risk diagnosis − 2.1766 × low-risk diagnosis − 1.7928, where each pupil, ventilation, elective, recovery and diagnosis indicator is 1 when present and 0 otherwise (Straney 2013, Table 3, p677). The three diagnosis tiers are ONE variable, not three: when conditions from more than one tier are present only the highest applies — very high-risk, then high-risk, then low-risk — and the others contribute nothing. SBP is in mmHg (unknown → 120; cardiac arrest → 0; shocked with an unmeasurable BP → 30) and enters both linearly and as SBP² ÷ 1000, which together are U-shaped with a minimum near 125.6 mmHg, so both low and high pressures raise the score. Base excess enters as its absolute value in mmol/L (unknown → 0); the oxygenation term is (FiO₂ × 100) ÷ PaO₂ with FiO₂ a fraction and PaO₂ in mmHg, or 0.23 when either is unmeasured (PIM3's substitute for a normal value, not PIM2's 0). Predicted mortality (probability) = 1 ÷ (1 + e^−logit). Both the logit and the probability (a value from 0 to 1) are reported, each to 4 decimal places; the derivation paper defines no severity bands.

PIM3 estimates the probability of death from data collected at first ICU contact. It is a unit-level case-mix / benchmarking tool — summed individual probabilities across a cohort give an expected death count, compared with observed deaths as a Standardised Mortality Ratio (SMR = observed/expected) — and is NOT an individual bedside prediction. The derivation paper says so itself, verbatim: "These models are not intended for prognostic use on individual patients" (Straney 2013), a sentence that then goes on to describe the group-level uses the model IS for. Straney 2013 also defines no diagnostic cut-points and no risk bands, so this score shows none — and that absence is settled rather than outstanding. No paediatric mortality model publishes endorsed severity tiers; registries report unit-level standardised mortality ratios with funnel plots rather than per-patient bands, and calibration papers use predicted-probability intervals only to test goodness of fit. Cutting a continuous prediction into categories is separately argued against on statistical grounds (Altman & Royston, BMJ 2006;332:1080), so the probability is presented whole. AGE RANGE — THE PAPER CONTRADICTS ITSELF. The abstract describes admissions of children younger than 18 at admission; the inclusion criteria in Methods state younger than 16. Read it as under 16, which is how the field reads it, and note that the discrepancy is in the source rather than resolved here: the Korean validation extended the model to under-18s precisely because the developmental data covered under-16s. DIAGNOSIS TIERS ARE ONE VARIABLE. Very high-risk, high-risk and low-risk are resolved to a single term by precedence — highest wins — and are never added together. This is a change from PIM2, where a high-risk and a low-risk condition could both count, and it is the commonest porting defect: on the paper's own worked example (p681) counting the high-risk term alongside the very high-risk one gives 72.34% instead of 47.22%. Each list is complete as published (5 very high-risk, 5 high-risk, 6 low-risk) and applies only to the MAIN reason for admission. MISSING-DATA CONVENTIONS ARE LOAD-BEARING: unknown systolic BP substitutes 120 mmHg, unknown base excess contributes 0, and an unmeasured FiO₂/PaO₂ sets that term to PIM3's normal-value substitute of 0.23 (a correction — PIM2 used 0). This path is the ordinary path, not an edge case: PaO₂ was missing for 55.8% and FiO₂ for 41.1% of the derivation cohort, and a US single-centre series reported base excess missing in 97.2%. SBP coding: cardiac arrest at admission → enter 0; shocked with an unmeasurable BP → enter 30; unknown → leave blank (120). The two SBP paired terms are U-shaped with a minimum near 125.6 mmHg, so SBP 0 adds about 2.70 to the logit relative to the unknown default — that is how the arrest case acquires its weight. Use the FIRST value of each variable from first face-to-face ICU-team contact up to 1 hour after ICU arrival (may include ED or retrieval data), not the worst. LIMITATIONS. Newborns are systematically over-predicted: they sit physiologically well below the 125.6 mmHg SBP nadir, so the blood-pressure terms inflate their score (observed for both PIM2 and PIM3 in the Italian validation). Haemato-oncology admissions are under-predicted, and badly: discrimination fell to c-index 0.66 against 0.74–0.83 in other subgroups, with observed mortality 18.73% against 7.13% predicted (Lee 2017). Neurological admissions were under-predicted in the derivation cohort itself — SMR 1.32 (1.16–1.50), the only diagnostic group significantly off in the original data. Calibration travels far worse than discrimination: Italy AUC 0.88 / SMR 0.98 (Hosmer-Lemeshow p = 0.21, good); Argentina AUC 0.83 / SMR 1.3 (p < 0.001); South Africa AUC 0.81 / SMR 1.28 (p < 0.001) with the HIGHEST SMR (6.67) in the lowest-risk decile. THE GULF NOW HAS ITS OWN EVIDENCE RATHER THAN A PROXY. Dubai (Malhotra 2019, single centre, n = 583, 46 deaths, 7.9%): AUC 0.78 (95% CI 0.69–0.87) with an overall SMR of 0.53, meaning the model over-predicted deaths for the unit as a whole, alongside SMR 2.1 in the SEPSIS subgroup — under-prediction inside an over-predicting unit. Those three are the robust findings from that paper, and the sepsis one is both uncontradicted and the clinically consequential one. WHAT THE SAME PAPER SAYS ABOUT LOW PREDICTED PROBABILITIES IS UNSTABLE, AND BOTH HALVES ARE RECORDED HERE BECAUSE EITHER ALONE MISLEADS. Cut finely, it reports SMR 2.67 among children whose predicted probability was 1–5% — severe under-prediction. Cut coarsely at a predicted probability of 14.3%, it reports SMR 0.33 below the cut against 0.72 above it — over-prediction across that same low range. Two stratifications of one cohort of 583 pointing in opposite directions in the same region of the scale is a subgroup instability, not a direction, so no claim about the low end is made here from it. Riyadh (Alkhalifah 2022, n = 3396, under 14 years) concluded that the mortality models it evaluated had sufficient discrimination and poor calibration, with the worst calibration AND the worst discrimination in infants under 12 months; the per-model figure recorded here from it is PRISM III's (AUC 0.87 in the 60–120-month band), and no PIM3-specific statistic from that study is claimed. READ THE SHAPE, NOT THE HEADLINE: discrimination travels between populations, calibration frequently does not, and the regional under-prediction this platform will state is the one that survives its own paper — SEPSIS, by a factor of about 2. A unit-level SMR below 1 does not make the model safe on the individual admissions inside it; it can conceal a doubling of risk in a subgroup. Recalibrate and monitor locally before any comparative interpretation. South Africa remains the only MULTICENTRE evaluation in a resource-varied setting, and is still the comparator for that specific question. HIV infection and admission after liver transplant were dropped from the model as non-predictive in the derivation population, but remain associated with higher mortality in the Argentine setting. COEFFICIENT SET. This is the published international model (Straney 2013). ANZICS also publishes regional recalibrations — PIM3-anz13 and PIM3-anz15 — whose coefficients are entirely different (anz13 pupils 4.371172, intercept −2.299542); do not mix them with these. REGISTRY CODE NUMBERS ARE NOT IMPLEMENTED, DELIBERATELY. The ANZPIC registry carries its own numbering for the three tiers, and it DIVERGES from the paper's: registry high-risk code 5 is septic shock (collected but not used by PIM3) while necrotising enterocolitis moves to code 6, and the registry adds very-high-risk combination codes 7 and 8 that a naive 1–5 membership test would silently drop. This calculator consumes no registry data, so it maps no codes; anyone who later ingests ANZPIC records needs two explicitly labelled mappers, not one. The registry numbering is documented in the ANZICS 'PIM2 & PIM3 for the ANZPIC Registry' Information Booklet, whose published URL returns HTTP 404 (re-verified 2026-08-02). THE TRACHEOSTOMY RULE IS A REGISTRY CONVENTION, AND THAT IS NOW ITS STATED PROVENANCE RATHER THAN A GAP. A tracheostomy with unassisted spontaneous breathing does not count as ventilation in the first hour. The source is the ANZPIC Registry's own data dictionary — 'PIM2 & PIM3 for the ANZPIC Registry — Information Booklet', version January 2019, whose text was obtained and read on 2026-08-03 and which states the exclusion in those terms. It is grey literature: no DOI, and its published anzics.org URL still returns HTTP 404, so no link is carried for it. NO PEER-REVIEWED SOURCE ADDRESSES THE EDGE CASE AT ALL — Straney 2013 Appendix 1 (p680) lists only what the criterion INCLUDES (invasive ventilation, mask or nasal CPAP, BiPAP, negative-pressure ventilation) and is silent on tracheostomy. Read the rule for what it is: how the derivation registry coded the variable, which is the best available answer and the one that keeps this calculator consistent with the data the model was fitted on, but not a finding from the paper.