Laboratory test-ordering patterns carry an early prognostic signal in electronic health records

Technical Summary: Laboratory Test-Ordering Patterns as Early Prognostic Signals in EHRs

Problem Statement
Electronic Health Records (EHRs) capture not only physiological measurements (e.g., laboratory results, vital signs) but also the continuous record of clinical decisions, workflows, and measurement practices. While routine laboratory data contain physiological information, their presence, absence, timing, and frequency also reflect clinician judgment, diagnostic uncertainty, and monitoring intensity. Previous research has utilized order data for decision support, clustering, or laboratory stewardship but has generally treated ordering as a target for utilization management rather than an independent prognostic signal. There is a gap in systematically evaluating whether physician laboratory test-ordering behavior, captured via metadata, provides an early and independent prognostic signal that can predict adverse outcomes before measured values are available or when those values appear reassuring.

Methodology
The study employed a retrospective cohort design using three distinct EHR datasets:

1. MIMIC-IV (Primary Cohort): 305,732 admissions from Beth Israel Deaconess Medical Center (2008–2019).
2. eICU-CRD (Zero-Shot External Validation): 108,324 ICU stays across 206 hospitals, used without retraining.
3. MC-MED (De Novo Replication): 118,385 Emergency Department visits from Stanford University Medical Center (2020–2022).

Feature Engineering:
Two parallel feature sets were constructed from the first 48 hours of admission:

• Measured Value Features: 184 features comprising summary statistics (mean, min, max, last value) for 35 laboratory tests and 11 vital-sign streams.
• Test-Ordering Pattern Features: 146 features derived exclusively from measurement-count metadata (no physiological values). These included total measurement counts, the fraction of 4-hour time bins with at least one measurement, linear temporal trends, and admission-level summaries of ordering intensity, diversity, breadth, and escalation.
• Static Baseline Features: 70 variables including demographics, comorbidities (Charlson Index), and medication classes.

Modeling Approach:
Light Gradient Boosting Machine (LightGBM) models were trained to predict four outcomes: in-hospital mortality, 30-day readmission, acute kidney injury (AKI), and sepsis. Three model types were compared:

1. Value Model: Static features + Measured values.
2. Ordering Model: Static features + Ordering patterns.
3. Combined Model: Static features + Measured values + Ordering patterns.

Key Contributions and Analyses

• Prognostic Comparison: Direct comparison of ordering-pattern models against value-based models to assess if metadata alone carries significant predictive power.
• Modality Attribution: Used Shapley values in a cooperative game framework (15 modality coalitions) to quantify the marginal contribution of ordering patterns relative to other data streams.
• "Deceptively Normal" Phenotype: Identified patients with low measured value abnormality (Q1) but high ordering intensity (Q4) to determine if ordering patterns flag high-risk patients missed by standard metrics.
• Interpretability and Robustness:
  • Compared ordering features against simplified SOFA scores.
  • Analyzed ordering trajectories (escalation vs. de-escalation) in patients with stable measured values.
  • Identified "sentinel" ordering patterns (specific tests with strong univariate prognostic signals).
  • Conducted subgroup analyses (race, insurance, admission timing) and dose-response analyses.
  • Performed instrumental-variable analysis (2SLS) to assess confounding by illness severity.
• Temporal and External Validation: Evaluated performance in the first 4 hours of admission (before many results are available) and tested generalizability across independent care settings (ICU and ED).

Results

• Predictive Performance: In MIMIC-IV, the ordering-pattern model (containing no measured values) achieved AUROCs approaching those of the value-based model.
  • Mortality: 0.905 (Ordering) vs. 0.932 (Values).
  • Sepsis: 0.912 (Ordering) vs. 0.924 (Values).
  • While measured values retained a small, statistically significant advantage, the ordering model recovered most of the discrimination using only metadata.
• Modality Contribution: Shapley value analysis ranked test-ordering patterns as the second or third most informative data stream across outcomes, surpassing vital signs in several cases.
• Deceptively Normal Phenotype: Patients with low value abnormality but high ordering intensity had significantly higher event rates (e.g., 16.5% sepsis vs. 0.5% in concordant normal patients). Propensity-score matching confirmed substantial risk elevation (OR 23.7 for sepsis) even after controlling for demographics and comorbidities.
• Interpretability: Ordering features added prognostic value beyond SOFA scores. A U-shaped mortality trajectory was observed in patients with stable values, where both increasing (concern for deterioration) and decreasing (therapeutic withdrawal/palliative care) ordering patterns signaled risk. Sentinel tests (e.g., frequent heart rate, lactate, creatinine) mapped to specific clinical concerns.
• Robustness: The signal remained strong in patients without documented comorbidities and was consistent across racial, insurance, and admission timing strata. Instrumental-variable analysis suggested the association was not solely driven by unmeasured severity.
• Early Prediction: Within the first 4 hours (when only ~26.6% of admissions had lab results), the ordering model approached the performance of the value model (Mortality AUROC 0.819 vs. 0.831).
• External Validation:
  • eICU-CRD: Zero-shot transfer showed performance attenuation (expected due to domain shift), but ordering patterns remained informative.
  • MC-MED: De novo training in the ED setting showed ordering patterns outperformed ED laboratory values for inpatient admission (AUROC 0.876 vs. 0.783) and were competitive for ICU admission.

Significance and Claims
The paper claims that physician laboratory test-ordering behavior is an early, reproducible, interpretable, and routinely available prognostic signal embedded in EHR data. The findings demonstrate that when and how often tests are ordered carries information beyond the measured values themselves. Specifically:

1. Ordering patterns can identify high-risk patients with "deceptively normal" physiological values who might otherwise be missed by early warning systems.
2. These patterns provide prognostic utility in the earliest hours of admission, before laboratory results are available.
3. The signal is robust across different care settings (ICU, ED, general inpatient) and demographic subgroups.

The authors conclude that if prospectively validated, ordering metadata could support earlier risk stratification and decision support with minimal additional burden, as the data is already generated during routine clinical care. They emphasize that ordering metadata should complement, not replace, physiological measurements and clinician judgment, serving as an additional layer capturing clinical concern and workflow intensity.