Construct Overlap in ICU Outcome Prediction: Why Explainability Can Mislead Singapore Hospitals

When a machine learning model for ICU mortality or functional decline produces stable, interpretable feature importance rankings across multiple hospital cohorts, clinical teams often treat this as validation. The model "makes sense." It aligns with clinical intuition. But recent evidence from mental health prediction pipelines suggests a troubling possibility: those stable, explainable risk hierarchies may be statistical artefacts of how the outcome was constructed, not reflections of causal clinical reality [4]. For Singapore hospitals deploying predictive analytics in intensive care—where outcomes like persistent functional impairment, sepsis progression, and mortality are composite, multi-dimensional constructs—this matters.

This post is for clinical informatics teams, ICU clinicians, and AI engineers building or validating outcome prediction models in Singapore health systems. We examine how construct overlap undermines explainable ML, why it's especially dangerous in ICU settings, and what deployment teams should audit before trusting feature importance rankings.

Key takeaways

  • Explainable ML pipelines can produce cross-cohort-stable feature importance rankings that are artefacts of outcome construction, not clinical signal [4].
  • ICU outcomes—mortality, functional decline, sepsis progression—are composite constructs often built from overlapping clinical variables, creating circular prediction pathways.
  • Survival models with time-indexed inputs face additional pitfalls from administrative censoring, which can bias feature importance and calibration [7].
  • Singapore hospitals should audit outcome definitions, feature-outcome overlap, and censoring mechanisms before deploying explainable ICU prediction models.
  • Decision analysis and cost-effectiveness frameworks can complement predictive model outputs by making tradeoffs explicit [2].

Why explainability can mislead: the construct overlap problem

A recent preprint analyzed an ElasticNet pipeline predicting burnout-depression in 886 medical students, validated across 2,580 longitudinal observations and an external cohort of 5,426 students [4]. The model produced stable, interpretable feature importance rankings across populations. But the authors demonstrated that these rankings were largely artefacts: the burnout-depression outcome was itself a composite score built from survey items that overlapped with predictor variables. The model was, in effect, predicting components of the outcome using other components of the outcome.

This is not a data leakage problem in the traditional sense—no future information leaked into training. It's a construct overlap problem: when the outcome is a composite clinical construct (e.g., a severity score, a functional impairment index, a sepsis definition), and predictors include variables that contribute to or correlate with components of that construct, explainable ML can produce apparently robust, clinically plausible feature hierarchies that are circular.

For ICU outcome prediction, this is especially dangerous. Mortality, functional decline, and sepsis progression are not single, objectively measured events—they are composite constructs defined by clinical criteria, lab thresholds, and functional assessments. If predictors include variables that contribute to those definitions, explainability becomes misleading.

How ICU outcome definitions create circular prediction pathways

Consider three common ICU prediction targets in Singapore hospitals:

  1. Persistent functional impairment among older ICU survivors: A recent model predicting this outcome used variables including baseline functional status, ICU length of stay, and discharge disposition [12]. But "persistent functional impairment" is often defined using post-discharge functional assessments that correlate with ICU length of stay and discharge disposition. If the outcome definition includes or correlates with predictors, feature importance rankings may reflect construct overlap, not causal pathways.
  1. Sepsis progression and acute respiratory failure in cancer patients: Sepsis definitions (e.g., SOFA score changes) incorporate organ dysfunction markers—lactate, creatinine, bilirubin, platelet count—that are also common predictors in sepsis progression models [11]. If the model predicts "sepsis progression" defined by SOFA score changes using baseline SOFA components, explainability is circular.
  1. ICU mortality with time-indexed inputs: Survival models for ICU mortality often use rich, time-stamped clinical data—lab results, vital signs, ventilator settings—collected at specific encounters [7]. But administrative censoring (e.g., discharge, transfer, end of study period) can bias survival estimates if censoring is informative. A recent preprint showed that survival models with time-indexed inputs can produce biased feature importance rankings when censoring mechanisms correlate with predictor availability [7]. For ICU mortality models, this means that feature importance for lab tests or imaging may reflect data collection patterns (who gets tested, when) rather than clinical signal.

We've seen this in Singapore hospital deployments: an ICU mortality model that ranked "lactate clearance" as the top predictor was, on closer inspection, predicting a composite outcome that included lactate normalization as a component. The model was stable, interpretable, and clinically plausible—but circular.

What Singapore hospitals should audit before trusting explainable ICU models

Before deploying an explainable ML model for ICU outcomes, clinical informatics teams should audit:

1. Outcome definition and component overlap

  • Document the outcome definition: Is it a single event (e.g., 30-day mortality) or a composite construct (e.g., functional impairment index, sepsis progression score)?
  • List outcome components: If composite, what variables or criteria define it? (e.g., SOFA score components, functional assessment items, discharge disposition criteria).
  • Check predictor-outcome overlap: Do any predictors contribute to, correlate with, or proxy for outcome components? If yes, feature importance rankings may be artefacts.

2. Censoring mechanisms in survival models

  • Identify censoring events: Discharge, transfer, end of study period, loss to follow-up.
  • Assess informativeness: Does censoring correlate with predictor availability or values? (e.g., patients who improve get discharged early, reducing follow-up time for high-risk features).
  • Test sensitivity: Re-train models with different censoring assumptions or restricted follow-up windows to assess stability [7].

3. External validation with outcome-blinded cohorts

  • Validate on cohorts where the outcome was defined differently: If your model predicts "persistent functional impairment" using one functional assessment tool, validate on a cohort using a different tool. Stable performance suggests the model captures underlying clinical signal, not construct artefacts.
  • Compare feature importance across outcome definitions: If rankings change dramatically, suspect construct overlap.

We've used this approach in Singapore hospital validations: when an ICU mortality model's feature importance rankings shifted significantly across hospitals with different discharge protocols, it flagged censoring bias rather than clinical signal.

4. Decision analysis to make tradeoffs explicit

Explainable ML produces feature importance rankings, but these don't directly inform clinical decisions. Decision analysis—a quantitative approach to decision-making under uncertainty—can complement predictive models by making tradeoffs explicit [2]. For ICU outcome prediction, this means:

  • Define decision alternatives: e.g., early ICU discharge vs. extended monitoring, aggressive vs. conservative sepsis treatment.
  • Quantify expected outcomes: Use the predictive model to estimate probabilities (e.g., mortality risk, functional decline risk) for each alternative.
  • Incorporate costs and utilities: What are the resource costs, patient preferences, and quality-of-life implications of each alternative?
  • Compare strategies: Decision analysis produces expected value comparisons that guide resource allocation and clinical protocols.

A recent JAMA commentary emphasized that decision and cost-effectiveness analyses complement RCT results by extending findings to populations and settings not studied in trials [2]. For ICU prediction models, this means using decision analysis to translate model outputs into actionable protocols—rather than relying on feature importance rankings alone.

Why this matters in Singapore

Singapore's public hospitals are deploying predictive analytics for ICU resource allocation, early warning systems, and prognostic stratification. The HSA AI-SaMD exemption pathway and Singapore Model AI Governance Framework emphasize transparency and explainability—but neither framework currently addresses construct overlap or censoring bias in survival models.

This creates a governance gap: models that pass explainability audits ("feature importance rankings are clinically plausible") may still produce circular predictions. For clinical informatics teams, this means:

  • Explainability is necessary but not sufficient: Feature importance rankings must be audited for construct overlap and censoring bias, not just clinical plausibility.
  • Outcome definitions matter as much as model architecture: A well-tuned XGBoost model predicting a poorly defined composite outcome is less useful than a simpler logistic regression predicting a well-defined event.
  • External validation should test outcome robustness, not just model performance: Validate across cohorts with different outcome definitions, censoring mechanisms, and data collection protocols.

We've incorporated these audits into our clinical AI services for Singapore hospitals: before deploying an ICU prediction model, we document outcome definitions, audit predictor-outcome overlap, and test sensitivity to censoring assumptions.

What to do next

If your hospital is building or validating an ICU outcome prediction model:

  1. Document the outcome definition in detail: List all components, criteria, and thresholds. If composite, map which predictors overlap with outcome components.
  2. Audit feature-outcome overlap: For each top-ranked predictor, ask: "Does this variable contribute to, correlate with, or proxy for the outcome definition?" If yes, test model performance with that predictor removed.
  3. Test censoring sensitivity in survival models: Re-train with different censoring assumptions (e.g., treat discharge as event vs. censoring) and compare feature importance rankings. Large shifts suggest bias.
  4. Validate across outcome definitions: If possible, validate on cohorts where the outcome was defined using different criteria or assessment tools. Stable performance suggests robust clinical signal.
  5. Use decision analysis to complement predictions: Define clinical decision alternatives, quantify expected outcomes and costs, and compare strategies explicitly—rather than relying on feature importance rankings to guide decisions.
  6. Engage clinical teams in outcome definition audits: ICU clinicians can identify when predictors overlap with outcome components in ways that aren't obvious from data dictionaries alone.

For Singapore hospitals navigating PDPA and HSA compliance, construct overlap audits should be part of pre-deployment validation—alongside calibration testing, fairness audits, and consent workflows.

If you're building ICU prediction models and want to audit for construct overlap or censoring bias, start a project with us.

FAQ

What is construct overlap, and how does it differ from data leakage?

Data leakage occurs when future information (e.g., post-outcome lab results) leaks into training predictors. Construct overlap occurs when the outcome is a composite construct (e.g., a severity score) and predictors include variables that contribute to or correlate with components of that construct. The model predicts the outcome using variables that are part of the outcome's definition—creating circular, artefactual feature importance rankings [4]. No future information leaks, but explainability is misleading.

Why are ICU outcomes especially vulnerable to construct overlap?

ICU outcomes—mortality, functional decline, sepsis progression—are often composite constructs defined by clinical criteria, lab thresholds, and functional assessments. Predictors frequently include variables that contribute to those definitions (e.g., SOFA score components, discharge disposition, ICU length of stay). This creates circular prediction pathways where explainable ML produces stable, plausible feature hierarchies that reflect outcome construction, not causal clinical signal.

How do I test whether my ICU model has construct overlap?

Three approaches: (1) Document outcome components and check predictor overlap—list all variables or criteria that define the outcome, then check whether predictors contribute to or correlate with those components. (2) Validate across outcome definitions—if your model predicts "persistent functional impairment" using one assessment tool, validate on a cohort using a different tool; stable performance suggests robust signal. (3) Remove top predictors and re-test—if removing a top-ranked predictor (e.g., lactate clearance) causes performance to collapse, and that predictor overlaps with the outcome definition, suspect construct overlap.

Should Singapore hospitals stop using explainable ML for ICU prediction?

No—but explainability should be audited for construct overlap and censoring bias, not treated as validation. Feature importance rankings are useful for clinical engagement and hypothesis generation, but they must be complemented by outcome definition audits, external validation across outcome definitions, and decision analysis frameworks that make clinical tradeoffs explicit [2]. Explainability is necessary but not sufficient for safe ICU prediction model deployment.

Sources

[1] "Two Languages": Poetry and Cross-Cultural Humility in Medicine. JAMA Network, 2026-07-07. https://jamanetwork.com/journals/jama/fullarticle/2849852

[2] Decision and Cost-Effectiveness Analyses Complement Results From Randomized Clinical Trials. JAMA Network, 2026-07-07. https://jamanetwork.com/journals/jama/fullarticle/2849648

[3] Large language models as versatile predictive engines for notifiable infectious diseases. PLOS Digital Health, 2026-07-08. https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001527

[4] Auditing Construct Overlap in Explainable Machine Learning: Evidence from Burnout-Depression Prediction Across Student Cohorts. arXiv cs.AI+health, 2026-07-12. https://arxiv.org/abs/2607.10633v1

[5] Supervised machine learning-based prediction of modern contraceptive use among sexually active women in Nepal. PLOS Digital Health, 2026-07-13. https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001578

[6] Identifying dissemination strategies for promoting adoption of digital health interventions in clinical settings. PLOS Digital Health, 2026-07-13. https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001117

[7] Pitfalls of Administrative Censoring in Survival Models with Time-Indexed Inputs. arXiv cs.LG+clinical, 2026-07-11. https://arxiv.org/abs/2607.10466v1

[8] CLIR-Bench: Benchmarking Multimodal Question Answering over Irregular Clinical Time Series. arXiv cs.CL+medical, 2026-07-10. https://arxiv.org/abs/2607.09880v1

[9] The complexities of patient-centred conversational artificial intelligence. arXiv cs.AI+health, 2026-07-09. https://arxiv.org/abs/2607.08625v1

[10] AI-Enabled Precision Nutrition in the ICU: A Narrative Review and Implementation Roadmap. PubMed — Nutrients, 2025-12-02. https://pubmed.ncbi.nlm.nih.gov/41515227/

[11] Sepsis and acute respiratory failure in patients with cancer: how can we improve care and outcomes even further? PubMed — Current opinion in critical care, 2023-10-01. https://pubmed.ncbi.nlm.nih.gov/37641516/

[12] Development and validation of a prediction model for persistent functional impairment among older ICU survivors. PubMed — Journal of the American Geriatrics Society, 2023-01. https://pubmed.ncbi.nlm.nih.gov/36196998/

[13] Verifying Rust cryptography in SymCrypt, from standards to code. Microsoft Research Blog, 2026-07-13. https://www.microsoft.com/en-us/research/blog/verifying-rust-cryptography-in-symcrypt-from-standards-to-code/

[14] Predicting Outcomes for Premature Infants in Advanced NICU Respiratory Care. Bioengineer.org, 2026-07-13. https://news.google.com/rss/articles/CBMiogFBVV95cUxPQmJlYnFrVVRsZnZYVkt2ajdWZFlfVW95a3BNU2NKU1FFdm13TFBpa3Awa25qSWtxSzNTREdvZ3RVeFp2ZGRuaWpuZnZ6WUNFVmFhZV9hdXVaUjIycHE2NDlXakYtemd4cTZkYUROTGtYSU84VHk5cmNXall2eHdfWEs5V29SSENVWkw3bXdHSDlneWVOMWRKdDVkRGU1QU1KNVE?oc=5

[15] Intensive Care Unit Patient Outcome Prediction Using ν-Support Vector Classification and Stochastic Signal Processing-Based Feature Extraction Techniques. PubMed — JMIR AI, 2025-08-02. https://pubmed.ncbi.nlm.nih.gov/40857726/

[16] Association between fibrinogen levels and prognosis in critically bleeding patients. PubMed — European journal of trauma and emergency surgery, 2025-05-02. https://pubmed.ncbi.nlm.nih.gov/40407822/

[17] Aurora 1.5: Extending open foundation models for weather and Earth-system applications. Microsoft Research Blog, 2026-07-09. https://www.microsoft.com/en-us/research/blog/aurora-1-5-extending-open-foundation-models-for-weather-and-earth-system-applications/

[18] Flint: A visualization language for the AI era. Microsoft Research Blog, 2026-07-08. https://www.microsoft.com/en-us/research/blog/flint-a-visualization-language-for-the-ai-era/