Case 02 · National health network
Multimodal clinical triage
A specialist-grade triage stack that reads imaging, vitals, notes, and labs in one pass — cutting decision latency from twelve minutes to forty-seven seconds with 99.2% concordance to specialist panels in shadow mode.
Context
Emergency and urgent pathways inside a national health network were bottlenecked by specialist availability. Junior clinicians made high-stakes routing decisions under time pressure, while imaging and labs arrived asynchronously into systems that did not speak to each other. Prior AI pilots scored radiology in isolation and ignored the note that said “patient declining.”
The problem, precisely
Accuracy without calibration is dangerous. A model that is “usually right” but confidently wrong on edge cases is worse than a slower human. The network needed multimodal reasoning and the ability to abstain — to escalate rather than invent certainty — with audit trails suitable for clinical governance and eventual regulatory review.
Design rule: clinicians must trust the system’s silence more than
its confidence. Abstention is a feature, not a failure.
What we built
A vision-language backbone fused imaging embeddings with structured vitals, lab time series, and clinical notes. Conformal prediction layers produced set-valued outputs with finite-sample coverage guarantees. Human override telemetry captured every disagreement for continuous evaluation.
Architecture highlights
- Multimodal encoder with modality dropout for missing inputs.
- Conformal prediction for calibrated abstention bands.
- Full provenance: which inputs drove which recommendation.
- Shadow deployment against specialist panel labels.
- Policy-as-code gates for contraindicated pathways.
The hard parts
Label quality varied wildly by site. We built a specialist review flywheel and refused to train on contested cases without adjudication. Distribution shift between urban trauma centers and rural clinics required site-aware calibration, not a single global threshold.
Change management mattered as much as ROC curves. We embedded clinical champions, measured override reasons, and published weekly “miss postmortems” that treated false confidence as a Sev-adjacent event even when no patient harm occurred.
Rollout
Shadow mode across three sites for eight weeks, then assisted mode where the system proposed and clinicians confirmed. Full assisted triage expanded only after concordance and override metrics cleared governance thresholds. Marketing and patient communications teams were looped in late — not to hype AI, but to set honest expectations about assisted care pathways.
Outcomes
Median decision latency moved from roughly twelve minutes to forty-seven seconds on the measured pathway. Concordance with specialist panels hit 99.2% in shadow evaluation. Zero Sev-1 incidents attributed to the system in the first 180 days of assisted operation.
What we’d repeat
Start with regulatory posture, not with a demo. Calibrate before you celebrate accuracy. And instrument disagreement as carefully as you instrument agreement — that is where safety lives.