Phoenix Sepsis Score

Total (0–13) = respiratory (0–3) + cardiovascular (0–6) + coagulation (0–2) + neurologic (0–2), each per JAMA 2024 Table 2 (PaO₂ and MAP in mmHg, FiO₂ a fraction). RESPIRATORY is cumulative and scored only when respiratory support is present. The PaO₂:FiO₂ (P/F) and SpO₂:FiO₂ (S/F) ratios are evaluated together — S/F only when SpO₂ ≤ 97% — and either can trigger a tier: 1 point for any support with P/F < 400 or S/F < 292, 1 more for invasive mechanical ventilation with P/F < 200 or S/F < 220, and 1 more for invasive ventilation with P/F < 100 or S/F < 148. Invasive ventilation can therefore reach 3 and non-invasive support cannot exceed 1. CARDIOVASCULAR sums three independent 0–2 sub-scores with no overall cap: vasoactive agents (0 → 0, 1 → 1, ≥ 2 → 2), lactate (< 5 → 0, 5 to < 11 → 1, ≥ 11 → 2 mmol/L), and age-banded MAP (2 if MAP < low, 1 if low ≤ MAP < high, 0 if MAP ≥ high, with half-open [low, high) thresholds by age band). COAGULATION adds 1 point each — platelets < 100 ×10³/µL, INR > 1.3, D-dimer > 2 mg/L FEU, fibrinogen < 100 mg/dL — capped at 2. NEUROLOGIC is hierarchical, not additive: bilaterally fixed pupils score 2 outright whatever the GCS; otherwise GCS ≤ 10 scores 1; otherwise 0. BOUNDARIES are compared continuously, matching the reference software rather than the integer bands printed for bedside use, at the four values where the two disagree: MAP 30.5 under 1 month scores 1 where the printed band reads 0; lactate 10.95 scores 1 where the printed band (5–10.9) leaves a gap; and a P/F of exactly 200 or an S/F of exactly 220 on invasive ventilation scores 1 where the printed 2-point bands (100–200, 148–220) read 2. MISSING INPUTS are each imputed at their own normal-end sentinel, independently of every other input — an uncomputable ratio as 500, GCS as 15, pupils as not fixed, ventilation and other support as absent, vasoactive agents 0, lactate 0, INR 0, D-dimer 0, platelets and fibrinogen normal — so a missing GCS never suppresses a fixed-pupils score. Sepsis is reported as 1 when infection is suspected/confirmed and total ≥ 2; septic shock as 1 when sepsis is met and the cardiovascular component ≥ 1.

Diagnostic criterion, not a graded severity ladder: sepsis = suspected/confirmed infection AND total ≥ 2; septic shock = sepsis AND cardiovascular component ≥ 1 (a cardiovascular point, not merely total ≥ 2). The vasoactive sub-score is a COUNT of distinct agents (0/1/≥2), not the VIS. NEUROLOGIC IS HIERARCHICAL, NOT ADDITIVE. Bilaterally fixed pupils score 2 outright, whatever the GCS; a GCS ≤ 10 with pupils not fixed scores 1; anything else scores 0. All four published tables print this row as three mutually exclusive columns, and the neurologic sub-score the task force selected came from PELOD-2, whose own neurologic component is a maximum-type rule over GCS and pupillary reactivity. The reference software writes the rule as an additive expression capped at 2, which is observationally identical because fixed pupils alone already saturate the cap — but an UNCAPPED sum would return 3 for fixed pupils at GCS 8 and is wrong. Until 2026-08-03 this calculator returned 1 for fixed pupils whenever the GCS was above 10 or was not entered; that was a defect, and it is the reason for the v2.0.0 bump. MISSING INPUTS ARE IMPUTED AT THEIR NORMAL END, EACH INDEPENDENTLY of every other input — the rule is per input, not 'the component scores 0'. A missing GCS imputes to 15 and missing pupil status to 'not fixed', separately, so a child with bilaterally fixed pupils and no GCS recorded still scores the full 2 neurologic points and still meets the total ≥ 2 threshold. Both JAMA tables state the general rule in their first table footnote: an unmeasured variable contributes no points. One honest qualification, so the claim is not overstated as the whole of the published method: in the derivation cohort missing values were first carried forward from physiologically appropriate earlier time windows, and only values still absent after that contributed zero. That is a property of how the development dataset was built and has no bearing on a single-timepoint bedside calculator, where only the second half applies. The consequence is that the total reflects DOCUMENTED dysfunction — a teaching UI should show a missing input as missing, not as normal. RESPIRATORY USES BOTH RATIOS. P/F and S/F are evaluated together and either can trigger a tier, exactly as the source expression writes it; a ratio that cannot be computed imputes to 500 and so triggers nothing. S/F is valid only when SpO₂ ≤ 97%. That gate has a consequence worth stating outright, because it runs against intuition: with no arterial gas, a child on invasive ventilation at FiO₂ 1.0 scores respiratory 3 at SpO₂ 97 but 0 at SpO₂ 98 — the ratio simply is not computable above 97, and the missing-input convention then contributes 0. A well-saturated child on maximal support can therefore fall below the total ≥ 2 sepsis threshold on respiratory grounds alone. Read that as ‘not measurable’, never as ‘not hypoxaemic’, and obtain a blood gas. THE SUPPORT GATE IS READ OFF THE TASK FORCE'S OWN PUBLISHED CODE, not inferred from the printed table. The published SQL builds a single flag — other respiratory support is true when FiO₂ exceeds 0.21 or the child is invasively ventilated — and then scores invasive ventilation × (the two lower-ratio tiers) plus other respiratory support × (the P/F < 400 or S/F < 292 tier). Three things follow structurally rather than by reading between the lines: a child on no support scores 0 however low the ratio goes; the 1-point tier needs only some oxygen or non-invasive support; and the 2- and 3-point tiers need invasive ventilation. pSOFA IS BUILT THE OTHER WAY UP, AND THE DIVERGENCE IS REAL. There, respiratory subscores 3 and 4 carry a support requirement and no lower band carries any, so an unsupported child is CAPPED at 2 rather than floored at 0. The same child, at the same ratio, off support, is Phoenix respiratory 0 and pSOFA respiratory 2 simultaneously — a two-point divergence produced entirely by where each table puts its support requirement, with no disagreement about the physiology. Both are correct implementations of their own instrument; neither is to be harmonised to the other. ONE OPERATIONAL DIFFERENCE FROM THAT SQL, and it is this calculator's choice. The SQL infers support from the data, counting any FiO₂ above room air as support, because it extracts from records that carry no support field. This calculator asks instead, and takes ‘no respiratory support’ at face value even when an FiO₂ above 0.21 has also been entered. That pairing is internally contradictory — a child on FiO₂ 0.5 is by definition receiving oxygen — and it is the one input combination where this score and the published extraction disagree: the extraction would award the 1-point tier, this calculator scores 0. Answer the support question from what the child is actually on, and the two agree everywhere. HIGH-FLOW NASAL CANNULA COUNTS HERE AND DOES NOT COUNT IN THE REGISTRIES. Phoenix explicitly includes high flow within other respiratory support. PICANet and ANZPIC both EXCLUDE high-flow nasal cannula from the mechanical-ventilation field they collect. So the same child on high flow is ‘supported’ for Phoenix and ‘not ventilated’ in both major paediatric registries — which matters the moment a Phoenix score is read alongside registry-derived case-mix, benchmarking or ventilation-day figures. What has NOT been retrieved, and must not be assumed: no cohort has quantified how much the CPAP-versus-high-flow choice, or counting high flow on one side and not the other, shifts the distribution of this score. The direction is obvious; the magnitude is unknown. BOUNDARY CONVENTION. The published tables are written for bedside use (integers, one-decimal lactate) while the software treats every input as continuous, and the two disagree at exactly four values. This calculator follows the software at all four, consistently: MAP 30.5 in a child under 1 month scores 1 where the printed band reads 0 (the same shape recurs at every age band's 0/1 edge); lactate 10.95 scores 1, where the printed 1-point band of 5–10.9 leaves it in a gap; P/F exactly 200 on invasive ventilation scores 1, where the printed 2-point band of 100–200 reads 2; S/F exactly 220 on invasive ventilation scores 1, where the printed 2-point band of 148–220 reads 2. There is no divergence at any coagulation or neurologic boundary, nor at P/F 100, S/F 148, MAP 17, MAP 30, lactate 5 or lactate 11. OUT-OF-RANGE INPUT IS REJECTED RATHER THAN COMPUTED, and that is the source behaviour rather than a local choice: the reference R package halts on a GCS outside 3–15 and on ventilation/support flags outside {0, 1}. The criteria exclude age 18 and over, so 216 months and above is rejected here; note that the reference R code's own `<= 216` bound admits exactly 18.0 years and diverges from the criteria it implements. The age ceiling is exclusive: every age below 216 months is accepted, including a fractional one such as 215.5, and 216.0 itself is not. SEDATION. The neurologic sub-score was pragmatically validated in sedated and non-sedated patients, with and without invasive ventilation. The derivation paper separately acknowledges that some organ-dysfunction measures may reflect iatrogenic effects or clinician choices rather than sepsis-related dysfunction, and names a reduced GCS under sedation as its example. A sedated child's neurologic point may therefore be measuring the sedation; that caveat comes from the authors, not from commentary. GENERALISABILITY. The higher-resource derivation data came exclusively from US tertiary paediatric centres. Some lower-resource sites did not record respiratory support or neurologic status even when it had been assessed, which constrained both the achievable score range and the score's measured performance at those sites. ‘REMOTE ORGAN DYSFUNCTION’ IS NOT PART OF THE CRITERIA and is deliberately not implemented: the derivation paper uses it as a descriptive subgroup (respiratory or neurologic dysfunction plus at least one point in a different organ system) to characterise a higher-mortality population, not as a diagnostic gate. Only the numeric total GCS (3–15) is consumed; the GCS response-descriptor wording is a separate instrument (Teasdale & Jennett) and is not reproduced or licensed by the Phoenix papers. The validated cohort was children < 18 years and excluded newborns during the birth hospitalization and infants < 37 weeks post-conceptional age — applying the score to those groups is outside the validated population, and age in months is not adjusted for prematurity. The 8-organ research extension Phoenix-8 is research-only, is not the diagnostic criterion and is out of scope. The MAP age bands are implemented as half-open [low, high) intervals (matching the reference package) so non-integer MAP scores exactly. PLAUSIBILITY BOUNDS ARE NO LONGER ALL OURS. Until this version every min/max on this score was labelled a guardrail of this platform's own invention. The Phoenix implementation notes publish a reasonable-value table covering these inputs, and five of the bounds declared here are identical to it: age months [0, 216) with an exclusive ceiling — the same half-open domain, which independently corroborates the exclusive-ceiling work done here before that table was read; FiO₂ 0.21–1.00; GCS 3–15; the vasoactive count as an integer 0–6; and SpO₂'s ceiling of 100 together with the published statement that a value above 97 is unusable for SpO₂:FiO₂, which is the gate applied in the calculation. Narrower here than published, and unchanged: PaO₂ 20–700 against a published range open above with a floor of 0; MAP 10–200 against 1–300; lactate 0.3–30 against 0–50; D-dimer 0.1–50 against 0–500; and platelets 5–1000, INR 0.8–10 and fibrinogen 30–800 against published ranges that set no upper bound at all. NO BOUND ON THIS SCORE MOVED in light of the published table, and none should: a form field refusing a typo is doing a different job from a data pipeline's outlier filter, and widening to match would only admit values no bedside entry should contain. The published table also bounds analytes this four-organ score does not consume (creatinine 0–50 mg/dL, total bilirubin 0–100 mg/dL, ALT, glucose 5–2000 mg/dL — the research extension's inputs) and gives systolic and diastolic pressures for sites deriving MAP from a cuff, which this calculator takes directly instead. One arithmetic check falls out of it: the published SpO₂:FiO₂ ceiling of about 461.9 is 97 ÷ 0.21, exactly what the ≤ 97% usability rule and the room-air FiO₂ floor entail together. WHAT AN OUT-OF-RANGE VALUE DOES THERE IS NOT UNIFORMLY WHAT IT DOES HERE, and the two halves of the reference behaviour differ from each other. The R package's own argument checks HALT — on a GCS outside 3–15 and on ventilation or support flags outside {0, 1} — which is where the rejection behaviour stated above comes from and is not a local choice. The reasonable-value table is the other half: it is data-preparation guidance for the analytes, where a value outside its range is set to null, and a null scores zero with no imputation. This calculator rejects in both cases, so it is the stricter of the two for every analyte, deliberately — a silently nulled lactate reads exactly like a normal one, which is the failure this platform's missing-data caution exists to prevent. Provenance, stated so it is not overclaimed: the reasonable-value table was read from the reference package's implementation-notes documentation; the machine-readable units file that page refers to could not be retrieved, so nothing here is cited to that file. A second published comparator from a different tradition, tighter than both: the PICANet Admission Dataset Definitions Manual v5.4 (November 2020) publishes collection ranges including PaO₂ 3–60 kPa (22–450 mmHg), lactate 0.2–15.0 mmol/L, systolic pressure 20–180 with a check above 200, and base excess −30 to +20. No newer PICANet manual exists — confirmed, not merely unlocated. AND THE SILENCE ELSEWHERE IS ITSELF THE FINDING: VPS, PC4 and PHIS publish NO public numeric plausibility or edit-check bounds. Confirmed negative — theirs sit behind login as proprietary data-quality rules, so there is nothing to go and find. That is the reason independent implementations of the same score diverge on what they will accept: for most inputs there is no public standard to converge on, and any two implementations' guardrails are their own until someone publishes.