Pediatric ICU Blood Pressure Targets and Predictive AI: Clinical Trial Uncertainty

Pediatric ICU Blood Pressure Targets and Predictive AI: Clinical Trial Uncertainty

A randomized clinical trial published this week in JAMA [1] challenges a foundational assumption in pediatric critical care: that tighter blood pressure control improves outcomes in hypotensive children. The trial found no significant difference in outcomes between permissive and conventional mean arterial pressure (MAP) targets—a result that should prompt every hospital deploying ICU predictive AI to audit how their models encode clinical uncertainty. This post is for clinical informatics teams, ICU directors, and AI engineers building or procuring predictive analytics systems in Singapore and Asia-Pacific hospitals.

Key takeaways

• A new JAMA trial [1][2] found no outcome difference between permissive (>5th percentile for age) and conventional MAP targets in critically ill hypotensive children, challenging standard practice assumptions.
• Predictive AI models trained on historical ICU data inherit treatment patterns that may reflect clinical equipoise, not established efficacy—models that penalize permissive targets may encode outdated assumptions.
• Singapore hospitals deploying ICU outcome prediction must distinguish between descriptive patterns (what clinicians did) and prescriptive guidance (what clinicians should do), especially when trial evidence is emerging.
• Governance frameworks should require model cards to document which clinical questions remain under active investigation, so users understand when predictions reflect consensus vs. controversy.
• Updating models after trial publication is necessary but insufficient; the real challenge is designing systems that flag clinical equipoise before definitive evidence arrives.

What the pediatric MAP trial reveals about predictive model assumptions

The JAMA trial [1][2] randomized 1,500 critically ill children with hypotension to either a permissive MAP target (above the 5th percentile for age) or a conventional target (above the 50th percentile). The primary outcome—a composite of mortality and major morbidity—showed no significant difference between groups. This is a classic equipoise result: two reasonable treatment strategies, no clear winner.

Now consider a predictive model trained on historical ICU data to forecast mortality or length of stay. If that model learned from an era when clinicians preferentially escalated vasopressors to achieve conventional targets, it may have encoded an association: permissive MAP → worse outcomes. But the trial suggests that association was confounded by unmeasured severity or by the fact that sicker patients received more aggressive treatment. The model didn't learn a causal relationship; it learned a treatment pattern that reflected clinical uncertainty.

This is not a hypothetical concern. We've reviewed ICU risk models in Singapore health systems where blood pressure features were among the top predictors of mortality, but the models had no mechanism to distinguish between hypotension as a marker of severity and hypotension as a modifiable risk factor. When such models are used to guide resource allocation or triage decisions, they risk penalizing patients managed with permissive strategies—even when, as the trial shows, those strategies are non-inferior.

Why Singapore ICU predictive AI must separate description from prescription

Most ICU predictive models are descriptive: they forecast outcomes based on observed patterns. But clinicians often interpret predictions as prescriptive: if the model says a patient with MAP at the 10th percentile has high mortality risk, the implicit message is "raise the MAP." This conflation is dangerous when the underlying evidence is equivocal.

Singapore hospitals face a specific challenge here. Our ICU populations are heterogeneous—term neonates, ex-preterm infants, children with congenital heart disease, oncology patients—and treatment norms vary across institutions. A model trained at one hospital may encode that site's blood pressure management culture, not a universal physiological truth. When that model is deployed elsewhere, it imports assumptions that may not hold.

We've seen this in practice: a hospital cluster piloting an ICU early warning system previously discussed found that the model's blood pressure thresholds were tighter than the receiving unit's actual practice. Clinicians began second-guessing their permissive approach, not because of new evidence, but because the model's training data reflected a different treatment philosophy. The trial published this week [1][2] suggests those clinicians were right to be skeptical.

The fix is not to remove blood pressure from models—it's a legitimate severity marker—but to annotate predictions with uncertainty flags when the feature space overlaps with areas of clinical equipoise. If a patient's MAP is in a range where trial evidence is inconclusive, the model should say so, rather than confidently predicting an outcome based on historical treatment patterns.

How clinical trial results should trigger model audits

When a major trial like this one [1][2] is published, it should trigger a structured audit of deployed predictive models. Here's a checklist we use with clinical AI services clients:

1. Feature importance review: Does the model use MAP, vasopressor dose, or related features as predictors? If so, what is their SHAP or permutation importance rank?
2. Subgroup analysis: Stratify predictions by MAP percentile. Do patients with permissive targets (5th–50th percentile) receive systematically higher risk scores than those with conventional targets, holding other features constant?
3. Training data audit: What was the distribution of MAP targets in the training cohort? If the data came from an era or site with aggressive blood pressure management, the model may have learned to penalize permissive strategies.
4. Outcome correlation: In the validation cohort, is there a causal pathway from MAP to the outcome, or is MAP a proxy for severity? If the latter, the model should not be used to guide MAP management.
5. Model card update: Document that MAP target selection remains an area of clinical equipoise, and that predictions should not be interpreted as endorsing a specific target range.

This audit is not a one-time exercise. As trial evidence accumulates—on permissive MAP, on restrictive fluid strategies, on early vs. late vasopressor initiation—models must be re-evaluated. The alternative is to deploy systems that encode yesterday's assumptions into tomorrow's care.

Why this matters in Singapore and Asia

Singapore's Health Sciences Authority (HSA) and the Model AI Governance Framework emphasize that AI systems must be "fit for purpose" and that developers must document known limitations. A predictive model that penalizes permissive MAP targets without disclosing the underlying clinical uncertainty fails this standard. It's not that the model is wrong—it may accurately reflect historical outcomes—but it's incomplete: it doesn't tell users that the treatment pattern it learned is under active investigation.

This matters acutely in pediatric ICUs, where Singapore hospitals often manage rare conditions (e.g., post-cardiac surgery, dengue shock) with limited local trial data. If we import models trained on Western cohorts, we inherit their treatment norms and their clinical controversies. The permissive MAP trial [1][2] was conducted across multiple countries, but the patient mix, vasopressor choices, and co-interventions may differ from Singapore practice. A model that worked in the trial sites may not generalize here—not because of population differences, but because of practice differences.

Asia-Pacific hospitals deploying ICU predictive AI should also consider the broader implication: many "standard" ICU interventions—from glycemic targets to transfusion thresholds to ventilator settings—have been revised by recent trials. If our models are trained on data from before those trials, they encode outdated practice. This is why clinical analytics platform engineering must include a governance layer that tracks which clinical questions are settled and which are not.

The case for equipoise-aware predictive systems

What would an equipoise-aware ICU prediction system look like? Here's a sketch:

• Ontology of clinical uncertainty: Maintain a structured list of interventions and parameters (MAP targets, fluid strategies, antibiotic timing) where trial evidence is inconclusive or evolving. Link this ontology to model features.
• Prediction annotations: When a model uses a feature in the equipoise list, flag the prediction with a note: "This forecast incorporates blood pressure data; current trial evidence does not establish a single optimal MAP target for this population."
• Counterfactual explanations: For patients in equipoise zones, generate counterfactual predictions: "If MAP were managed permissively (10th percentile), predicted mortality would be X; if managed conventionally (50th percentile), predicted mortality would be Y." If X ≈ Y, the model is signaling equipoise.
• Trial-triggered retraining: When a major trial is published, automatically re-evaluate model performance on subgroups defined by the trial's intervention. If the model's predictions are inconsistent with trial findings, retrain or retire the model.

This is not science fiction. The infrastructure exists—ontologies like SNOMED CT and RxNorm, causal inference libraries like DoWhy, and MLOps platforms that support automated retraining. What's missing is the institutional commitment to treat clinical uncertainty as a first-class modeling concern, not an afterthought.

We've prototyped elements of this approach with institutional partners in Singapore, particularly around agentic clinical workflows where AI systems must reason about when to defer to human judgment. The pediatric MAP trial [1][2] is a reminder that "defer to human judgment" is not a fallback—it's the correct response when evidence is equivocal.

What to do next

• Audit deployed ICU models: If your hospital uses predictive analytics for ICU triage, mortality forecasting, or resource allocation, audit whether blood pressure features are used and how they influence predictions. Request SHAP plots or feature importance rankings from your vendor.
• Update model cards: Document which clinical interventions in your model's feature space are under active trial investigation. Make this information visible to end users, not buried in technical documentation.
• Establish a trial-monitoring process: Assign a clinical informatics lead to track major ICU trials (JAMA, NEJM, Critical Care Medicine) and trigger model audits when results are published. This should be a standing agenda item for your AI governance committee.
• Pilot equipoise annotations: For one high-stakes model (e.g., sepsis early warning), add a simple annotation layer that flags predictions involving parameters with uncertain evidence. Measure whether this changes clinician trust or decision-making.
• Engage with InsytAI services: If you're building or procuring ICU predictive AI and need help designing governance processes that account for clinical uncertainty, start a project with our team. We've worked with Singapore health systems to operationalize these principles.

FAQ

What is clinical equipoise and why does it matter for AI?

Clinical equipoise exists when expert clinicians genuinely disagree about which treatment is best, or when trial evidence has not yet established superiority. It matters for AI because models trained on historical data may encode one side of that disagreement as "correct," even though the evidence doesn't support that conclusion. The permissive MAP trial [1][2] is a clear example: models trained to penalize permissive targets would be encoding a preference, not a fact.

Should we stop using blood pressure in ICU prediction models?

No. Blood pressure is a legitimate marker of physiological instability and should remain in models. The issue is not the feature itself, but how predictions are interpreted. Models should distinguish between "low MAP is associated with worse outcomes in our training data" (descriptive) and "raising MAP will improve outcomes" (prescriptive). The latter requires causal evidence, which the recent trial [1][2] suggests is lacking for pediatric MAP targets.

How often should we retrain ICU models after new trials?

There's no universal answer, but a reasonable heuristic is: retrain when a trial challenges a feature that ranks in your model's top 10 predictors, or when the trial's population overlaps substantially with your patient cohort. For the pediatric MAP trial [1][2], any model using MAP percentiles or vasopressor doses in critically ill children should be audited. Retraining may not be necessary if the model's predictions are already consistent with equipoise, but you won't know until you check.

What if our vendor won't disclose model internals?

This is a governance red flag. Singapore's Model AI Governance Framework and HSA's software-as-medical-device guidance both emphasize transparency. If a vendor cannot explain which features drive predictions or how the model would respond to new trial evidence, that system is not fit for deployment in a governed clinical environment. Insist on model cards, feature importance documentation, and a process for post-market surveillance. If the vendor refuses, consider alternative solutions or contact us for procurement guidance.

Sources

[1] Permissive Blood Pressure Targets in Pediatric Critical Care. JAMA Network, 2026-10-27. https://jamanetwork.com/journals/jama/fullarticle/2854755

[2] Permissive Blood Pressure Targets in Pediatric Critical Care Research Summary. JAMA Network, 2026-10-27. https://jamanetwork.com/journals/jama/fullarticle/2854754

[3] Licensure Framework for Autonomous Clinical Artificial Intelligence. JAMA Network, 2026-10-27. https://jamanetwork.com/journals/jama/fullarticle/2854673

[4] M4Bench: Evaluating Procedural Specification for Clinical EHR Derivation Agents. arXiv cs.AI+health, 2026-10-27. https://arxiv.org/abs/2610.37459v1

[5] AI-Enabled Precision Nutrition in the ICU: A Narrative Review and Implementation Roadmap. PubMed — Nutrients, 2025 Dec 2. https://pubmed.ncbi.nlm.nih.gov/41515227/

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