CELTIC Cellulitis Tiered Interpretable Classifier

Cellulitis or a noninfectious mimic, in a hospitalized adult with a red, warm, swollen limb.

Research use only, not yet validated for clinical decisions. Everything runs in your browser. Nothing you enter is sent, stored, or logged.

You do not need every answer: the score reads from the first one, and anything left blank counts as no, as missing values did in the study. Rapid adds five findings once the first labs return. Standard and Comprehensive read the electronic record automatically, so there is nothing to enter by hand.

Finding Answer

Probability of cellulitis

 

Answer what you know. The score reads from the first answer.

Rule-out below 30%. Rule-in at 65% or above. Both cut points were fixed before evaluation.

What moved it log-odds, largest first

Each bar is that finding's coefficient. Right of the line favors cellulitis, left favors a mimic.

    About CELTIC

    Every tier is logistic regression. The coefficients on this page are the model.

    Use it for

    A hospitalized adult with a red, warm, swollen limb, as a transparent estimate to set beside your examination.

    Do not use it for

    Children, outpatients, an obvious alternative diagnosis, necrotizing infection, or sepsis. It does not replace the examination, an ultrasound for suspected clot, or a dermatology consult. Never withhold antibiotics from an unstable patient because of it.

    Limits

    The reference standard was the coded discharge diagnosis, not expert adjudication. Discrimination was lowest at 70 years or older (0.70), where a mimic is most likely, so look twice at a rule-in in that group. No tier uses a culture result.

    Trained on

    Patients
    24 852 adults, one index admission each
    Hospital
    Beth Israel Deaconess, 2008 to 2022 (MIMIC-IV v3.1)
    Cellulitis or erysipelas
    47.4%
    Mimics
    18 pseudocellulitis conditions, such as deep vein thrombosis, venous stasis, contact dermatitis, gout
    Laterality
    Read from 17 024 discharge summaries by a language model on local hardware

    Tested on

    AUROCCELTICComplete ALT-70
    Held-out fifth, Bedside tier 3446 with a processed note, 45.4% cellulitis0.72 0.70 to 0.740.60 0.59 to 0.62
    Other hospitals eICU-CRD, 2536 patients, structured version0.690.54
    Emergency department 3314 patients, no examination finding available0.560.52
    Paired difference
    +0.12 (0.10 to 0.14), P < .001, same held-out patients
    Other tiers
    Rapid 0.73, Standard 0.75, Comprehensive 0.76
    Calibration
    Slope 1.00, intercept −0.01

    Research use only. This calculator accompanies a manuscript under review and has not been validated prospectively. It is for research and education. The authors accept no responsibility for clinical decisions made with it.

    The arithmetic, worked by hand

    logit = −1.076 + 1.530·unilateral + 0.187·leukocytosis − 0.654·age ≥ 70 − 0.562·malignancy + 0.399·diabetes
    probability = 1 / (1 + e−logit)

    Example: unilateral involvement, diabetes, age 74, no leukocytosis, no malignancy.

    logit = −1.076 + 1.530 + 0.399 − 0.654 = 0.199
    e−0.199 = 0.820
    probability = 1 / 1.820 = 0.549, shown as 55%, indeterminate

    Hand score: unilateral +3, diabetes +1, leukocytosis 0, age 70 or older −1, active malignancy −1. This patient totals +3 + 1 − 1 = +3. In the test set, 60% of patients with a hand score of +3 had cellulitis.

    The Rapid tier has the same form with 10 coefficients, printed in the source.

    Cite

    Wu E, Lee A, Youssef A, English J, Mason E, Xing EP, Pugliano-Mauro M. CELTIC, a tiered interpretable model to distinguish cellulitis from mimics in hospitalized adults. JAMA Dermatol. 2026; submitted. cellulitiscalc.com