KATE Sepsis: Early ED Triage Risk Score

• KATE Sepsis is a machine-learning risk score for early sepsis detection at ED triage using only pre-laboratory data.
• It employs a two-stage stacked classifier combining XGBoost and logistic regression to integrate structured vital signs with NLP-extracted clinical features.
• Evaluation on 512,949 adult encounters showed improved sensitivity (71.09%) and AUC (0.9423) compared to the standard SIRS screening protocol.

KATE Sepsis denotes a machine-learning risk score for the early detection of sepsis during emergency department triage using only information available before laboratory diagnostics. It was developed from retrospective, de-identified triage data collected across 16 hospitals and evaluated against a standard screening protocol based on SIRS with source of infection on 512,949 adult encounters. In that evaluation, KATE Sepsis reported substantially higher discrimination than the rule-based comparator while operating at a similar specificity (Ivanov et al., 2022). In the sepsis literature, the name “KATE” also refers to “Knowledge-Abstraction Temporal-pattern Extraction,” an ICU-focused temporal abstraction and interval-pattern mining framework; despite the shared acronym, that framework is methodologically distinct from the ED triage classifier (Sheetrit et al., 2017).

1. Clinical setting and problem formulation

Sepsis is described as a life-threatening condition with organ dysfunction and as a leading cause of death and critical illness worldwide. The central clinical motivation for KATE Sepsis is the observation that even a few hours of delay in treatment results in increased mortality. The model is therefore positioned at emergency department triage, before laboratory diagnostics, where earlier recognition could permit earlier lab analysis, antibiotic administration, and other sepsis treatment protocols (Ivanov et al., 2022).

The study asked whether a machine-learning-based risk score, computed solely from triage data, could substantially outperform the ubiquitous SIRS-based screening protocol for the early identification of sepsis, severe sepsis, and septic shock. This framing is important because the comparator was not another high-capacity statistical model but a widely used rule-based screen. A plausible implication is that the study was designed to test whether richer use of routinely available triage information can compensate, at least in part, for the absence of laboratory features.

The labeling framework in the ED study used an ED clinician-assigned sepsis diagnosis based on Sepsis-2 criteria. Future work was explicitly recommended to include retraining or re-evaluation under Sepsis-3 criteria. This distinction matters because reported performance is anchored to a specific clinical definition rather than to a definition-independent notion of physiologic deterioration.

2. Cohort construction and feature representation

The KATE Sepsis cohort consisted of retrospective, de-identified triage data from 16 hospitals spanning February 2015 to July 2021. The raw cohort contained 615,581 adult encounters for patients aged at least 18 years. After excluding pediatric cases and records missing four or more of the six vital signs, 512,949 encounters remained, of which 9,257 had an ED clinician-assigned sepsis diagnosis (Ivanov et al., 2022).

Input representation combined structured triage variables, categorical metadata, and features extracted from free text. The numerical features measured at triage were age, heart or pulse rate, respiratory rate, temperature, systolic and diastolic blood pressure, oxygen saturation, primary pain intensity, Glasgow Coma Scale score, and point-of-care blood glucose. The categorical features were sex, arrival mode, and immunization status. Arrival mode included categories such as walk-in, ambulance, wheelchair, stretcher, and law enforcement; immunization status was encoded as up to date $=1$ and not up to date $=0$ (Ivanov et al., 2022).

Unstructured fields were processed through a custom clinical-concept extraction pipeline. Chief complaint, past medical and family history, surgeries, triage treatments, and pre-arrival medications were parsed by sentence and word tokenization, normalization, POS-tagging, chunking, and matching noun- or verb-phrase permutations to UMLS concepts. Each extracted concept was represented by its unique CUl code. Across the 512,949 records, thousands of unique clinical terms were discovered, including approximately $7.5$ terms per “reason for visit” and approximately $0.3$ in “family history” (Ivanov et al., 2022).

Preprocessing also included removal of physiologically impossible outliers in numerical fields. Missing values in the final feature matrix were handled natively by XGBoost. This feature design situates KATE Sepsis between conventional structured-score systems and broader clinical NLP pipelines: it remains triage-only, but it does not restrict itself to vital signs alone.

3. Model architecture and training procedure

KATE Sepsis is a two-stage stacked classifier that combines gradient-boosted decision trees with a final logistic regression layer. In the first stage, XGBoost produces a probability of sepsis for each encounter. If $F_t(x)$ is the prediction from tree $t$, the raw logit is

$z_x = \sum_{t=1}^{T} F_t(x).$

XGBoost optimizes a regularized logistic objective with log-loss,

$$L = - \frac{1}{N} \sum_{i=1}^{N} \left[y_i \log p_i + (1-y_i)\log(1-p_i)\right],$$

where $p_i = \sigma(z_i)$ and $\sigma(z)=1/(1+e^{-z})$ (Ivanov et al., 2022).

In the second stage, the XGBoost-derived probability and a subset of clinical features are passed to a sparse logistic regression model, again using the logistic function

$=0$0

with cross-entropy loss. Class imbalance was addressed by specifying “balanced” class weights in XGBoost, and missing feature values were left to XGBoost’s internal default imputation (Ivanov et al., 2022).

Training used five-fold cross-validation on the full 512,949-record cohort. In each fold, $=0$1 of the data were used for training and $=0$2 for validation, with the folds cycled so that every record served once as held-out test data. Preprocessing within the training pipeline included physiologic outlier removal, text normalization and concept extraction for unstructured fields, and one-hot encoding for categorical variables. Hyperparameter selection and early stopping were managed within each fold to avoid overfitting. The architecture therefore combines nonlinear interactions learned by boosting with a sparse linear calibration or integration layer, although the source does not provide an ablation isolating the contribution of each stage.

4. Evaluation protocol and quantitative performance

The study evaluated discrimination with the area under the receiver-operator characteristic curve,

$=0$3

where $=0$4 is sensitivity and $=0$5specificity. Sensitivity was defined as $=0$6, and specificity as $=0$7. Confidence intervals were computed by nonparametric bootstrap over patient-level predictions within each test fold, typically using 1,000 bootstrap replicates and reporting the 2.5th and 97.5th percentiles (Ivanov et al., 2022).

On the full adult cohort, averaged over five cross-validation folds, KATE Sepsis achieved an AUC of $=0$8 with $=0$9 CI $7.5$0 to $7.5$1, sensitivity of $7.5$2 with $7.5$3 CI $7.5$4 to $7.5$5, and specificity of $7.5$6 with $7.5$7 CI $7.5$8 to $7.5$9. The standard SIRS-with-infection screen achieved an AUC of $0.3$0 with $0.3$1 CI $0.3$2 to $0.3$3, sensitivity of $0.3$4 with $0.3$5 CI $0.3$6 to $0.3$7, and specificity of $0.3$8 with $0.3$9 CI $F_t(x)$0 to $F_t(x)$1 (Ivanov et al., 2022).

Metric	KATE Sepsis	SIRS with infection
AUC	0.9423 (0.9401–0.9441)	0.6826 (0.6774–0.6878)
Sensitivity	71.09% (70.12–71.98%)	40.80% (39.71–41.86%)
Specificity	94.81% (94.75–94.87%)	95.72% (95.68–95.78%)
Severe sepsis sensitivity	77.67% (75.78–79.42%)	43.06% (41.00–45.87%)
Septic shock sensitivity	86.95% (84.20–88.81%)	40.00% (36.55–43.26%)

The reported sensitivity gain was approximately 30 percentage points at essentially identical specificity, with $F_t(x)$2. The study further reported that KATE Sepsis correctly identifies $F_t(x)$3 of the cases that SIRS detects, whereas SIRS captures only $F_t(x)$4 of the cases KATE Sepsis flags. The trade-off was described as a very minor reduction in specificity, less than $F_t(x)$5, in exchange for a much larger increase in true positives. Within the study’s clinical framing, this trade-off was treated as especially relevant because missed sepsis cases at triage can delay time-sensitive treatment.

5. Feature importance, interpretability, and comparison with rule-based screening

Figure 1 in the ED study ranks the top 20 features by relative gain in the XGBoost stage. The summary emphasizes that, whereas SIRS relies solely on temperature, heart rate, and respiratory rate, KATE Sepsis also exploits age and weight; systolic, diastolic, and mean arterial pressures; oxygen saturation and flow rate; arrival mode; Glasgow Coma Scale; point-of-care blood glucose; and extracted clinical terms indicating comorbidities such as diabetes mellitus and symptoms such as generalized weakness, disturbance of consciousness, and cough (Ivanov et al., 2022).

This broader feature scope is central to the model’s conceptual difference from SIRS. SIRS is a threshold-based physiologic screen, whereas KATE Sepsis integrates structured measurements, triage logistics, and clinical language. The study explicitly argues that this broader feature set enables detection of septic physiology and risk factors that SIRS cannot capture at triage. A plausible implication is that part of the model’s gain derives from combining direct physiologic signals with proxies for acuity and pretest probability, such as arrival mode and free-text symptom descriptions.

Interpretability in KATE Sepsis is therefore partial rather than rule-exact. Relative gain rankings provide feature-level visibility, but the final prediction arises from a stacked ensemble rather than from a short deterministic checklist. The study does not claim full mechanistic interpretability; instead, it demonstrates that the predictor is not confined to the canonical SIRS variables. For readers accustomed to ML-based CDS systems, the model occupies an intermediate position: more expressive than rule-based screening, but still amenable to global importance analysis.

6. Relation to the ICU KATE framework and broader research context

A separate line of sepsis research uses the acronym KATE to denote “Knowledge-Abstraction Temporal-pattern Extraction,” an ICU-oriented temporal abstraction framework rather than the ED triage classifier. In that work, multivariate time-stamped ICU data are converted into clinically meaningful intervals by expert-defined thresholds $F_t(x)$6, $F_t(x)$7, and trend threshold $F_t(x)$8, producing qualitative states such as LOW, NORMAL, and HIGH, and gradients such as INCREASING, DECREASING, and STABLE. Adjacent time points with the same label are merged into maximal intervals, and KarmaLego is then used to mine frequent Time-Interval Relation Patterns over 26 sepsis-related concepts in 2,560 MIMIC-III ICU patients (Sheetrit et al., 2017).

The ICU KATE study reported that the distribution of temporal patterns with support above $F_t(x)$9 differed significantly between septic and non-septic patients in both 12-hour and 6-hour windows before sepsis onset. For the 12-hour window, 26,968 patterns reached the support threshold, including 6,168 unique to septic patients, 6,416 unique to non-septic patients, and 14,384 shared. For the 6-hour window, 22,422 patterns reached the threshold, including 5,356 unique to septic patients, 5,349 unique to non-septic patients, and 11,717 shared. The two-sample Kolmogorov–Smirnov tests yielded $t$0 for the 12-hour window and $t$1 for the 6-hour window, both greater than $t$2, with $t$3. Two-proportion tests further found that $t$4 of shared TIRPs in the 12-hour window and $t$5 in the 6-hour window had significantly different supports between classes (Sheetrit et al., 2017).

The relationship between these two KATE lines of work is primarily nominal rather than architectural. The ED KATE Sepsis system is a triage-only stacked classifier built from single-encounter data available prior to laboratory diagnostics, whereas the ICU KATE framework is an expert-driven temporal abstraction and interval-pattern mining approach over 6- to 12-hour windows. This suggests that “KATE Sepsis” spans at least two distinct methodological traditions in sepsis informatics: one centered on high-dimensional supervised risk scoring at the point of triage, and another centered on interpretable temporal pattern discovery in the ICU.

7. Clinical implications, limitations, and prospective directions

The ED KATE Sepsis study frames early recognition as operationally consequential because every hour of delay in antibiotic administration increases mortality. By identifying more than $t$6 of sepsis presentations at triage with more than $t$7 specificity, the system was described as potentially capable of triggering laboratory work-ups and imaging more promptly, accelerating antibiotic delivery toward the “golden hour” recommended by Surviving Sepsis Campaign guidelines, and potentially reducing morbidity, mortality, length of stay, and overall cost of care (Ivanov et al., 2022).

These clinical implications remain prospective rather than demonstrated outcomes in the reported study. The authors recommend prospective clinical validation, including real-time impact on time to antibiotics, ICU transfers, readmission rates, and in-hospital mortality. They also recommend retraining or re-evaluation under Sepsis-3 criteria. This emphasis on prospective validation is significant because the published analysis is retrospective and uses cross-validation within an historical cohort rather than live deployment.

Several boundaries of interpretation follow directly from the study design. The model was evaluated on adult encounters only; it excluded pediatric cases and records missing four or more of the six vital signs. Performance was measured against clinician-assigned ED diagnoses based on Sepsis-2 criteria. The source does not report prospective workflow effects, calibration under deployment shift, or inter-hospital transportability beyond the fact that the training and evaluation data came from 16 hospitals. For that reason, the most defensible characterization is that KATE Sepsis demonstrates strong retrospective triage-stage discrimination relative to SIRS, while its effect on downstream treatment timing and clinical outcomes remains a stated target for future research rather than an established result.