JMIR Publications

ABSTRACT

Background:

Acute hypotensive episodes are common in ICU care and often require rapid interpretation of severity and trajectory. Conventional rule-based clinical decision support can standardize thresholds, but it does not usually show how similar prior patients were managed or how they subsequently evolved.

Objective:

To develop and assess the feasibility of trend-aware patient similarity retrieval for acute hypotensive episodes and summarize retrieved-neighbor early actions and post-index trajectories.

Methods:

We performed a retrospective prototype study using adult ICU stays from MIMIC-IV. The primary cohort used one index event per ICU stay, defined as the first sustained hypotensive episode with 2 mean arterial pressure (MAP) readings below 65 mmHg within 60 minutes; a sensitivity cohort used the first MAP below 65 mmHg after ICU admission when the sustained criterion was not observed. Pre-index patient snapshots combined demographics, comorbidities, vitals, labs, pre-index interventions, and temporal trend features. Similar-patient retrieval used structured similarity after standard preprocessing and exclusion of the query case itself. Feasibility assessment focused on subgroup heterogeneity, retrieval validity versus random neighbors, exploratory clinical alignment and coherence, and comparison of latest-only versus trend-aware retrieval. The dashboard was not clinically tested, and no causal or treatment-effect claims were made.

Results:

The derived dataset contained 65,770 index events from 48,113 unique subjects and 65,770 ICU stays. The primary sustained cohort included 48,045 index events, and the fallback sensitivity cohort included 17,725. MAP trend features were available in 89.2% of snapshots, whereas lactate trend features were available in 30.7%. In a sampled retrieval evaluation of 50 query cases with top-k=20, trend-aware retrieval reduced MAP trajectory mean absolute difference from 14.12 to 12.51; aggregate neighbor-versus-random improvement was 34.2% (95% CI [0.294, 0.386]). Vasopressor action agreement was 87.1% (95% CI [0.807, 0.928]) for trend-aware retrieval versus 82.9% for random neighbors. Exploratory coherence analyses were directionally supportive but statistically inconclusive. In descriptive heterogeneity analyses, fluid exposure increased from 57.1% in lactate <2 mmol/L to 73.1% in lactate >4 mmol/L, while ICU mortality increased from 10.0% to 34.6%.

Conclusions:

This retrospective prototype suggests that trend-aware patient similarity retrieval may provide contextual summaries around acute hypotensive episodes by surfacing common early actions and divergent post-index trajectories among similar cases. Further clinician evaluation and prospective validation are required before clinical use.