Supporting Research
1. The unmet clinical need
1.1 Scale of the problem
Critical care medicine — operating rooms and intensive care units — represents the highest-acuity, highest-cost, and highest-information-density segment of hospital care: according to Halpern and Pastores, critical care accounts for 13.2–13.4% of U.S. hospital costs and 4.1% of national health expenditure (Crit Care Med, 2010). Order of magnitude for the bedside data stream:
- about 187 alarms per bed per day — roughly 68,000 per year per bed, and about 800,000 per year for a 12-bed unit (Drew et al., PLoS ONE, 2014)
- the share of false or clinically insignificant alarms — 72–99% depending on alarm type; for arrhythmia alarms, 88.8% false-positive has been measured (Drew et al., 2014; Sendelbach and Funk, AACN Adv Crit Care, 2013)
- 2–4 TB of raw monitoring data per year for a 12-bed unit (expert estimate)
The consequence: genuinely actionable alarms — those indicating real deterioration — are lost in noise. Delayed recognition of deterioration is a recognised problem of monitoring organisation: aggregate early-warning systems, such as the Royal College of Physicians’ NEWS2, respond to it.
1.2 Why existing tools fail
Threshold-based alarms — the dominant ICU monitoring paradigm — generate alarms when a single parameter crosses a static threshold. They do not model the patient’s trajectory, context, or the correlation between parameters. Result: high false-positive rate, high alarm fatigue.
AI-based CDS — the increasingly popular alternative — improves sensitivity but introduces new failure modes: model drift, lack of auditability, regulatory complexity, and automation bias. A high-confidence AI alert tends to suppress clinical skepticism precisely in the cases where skepticism is most warranted: systematic reviews document this effect and show that its magnitude depends on mediators — workload, trust in the system, and prompt placement on screen (Goddard, Roudsari, Wyatt, JAMIA, 2012; Lyell and Coiera, JAMIA, 2017).
The deterministic supervisory control system is the third path: rule-based analysis that is transparent, reproducible, and auditable, with a formal governance process that keeps clinical expertise central to every rule.
2. Market context and geography
2.1 Target market prioritisation
Geographic entry prioritisation is based on three criteria: regulatory pathway accessibility, clinical institution partner availability, and unmet need density.
Sub-Saharan Africa. South Africa (SAHPRA), Botswana (BoMRA): regulatory pathways are accessible for CE-marked or innovation-track medical software; critical care capacity is constrained; clinical institutions are actively seeking CDS solutions that do not require full ICU staffing. Primary source — SAHPRA · Primary source — BoMRA
Latin America. Brazil (ANVISA RDC 657/2022): large private hospital segment with procurement authority; RDC 657/2022 provides a Class II software pathway. Primary source — ANVISA
Southeast Asia. Malaysia (NPRA), Singapore (HSA/NMRC): advanced clinical infrastructure, regulatory frameworks compatible with EU MDR precedent, active interest in AI-adjacent CDS procurement.
East Africa. Kenya (PPB): emerging critical care infrastructure, WHO emergency care program alignment. Primary source — PPB
2.2 Volume estimates
Based on published national audits and surveys, the combined critical care bed count across priority-1 and priority-2 markets is at least 46,500 beds: South Africa — about 4,700, Botswana — a few dozen in the public sector, Brazil — 42,000–44,000, Kenya — about 500–600. General context — the World Bank indicator “Hospital beds per 1,000 people” → and the WHO resolution WHA72.16 on emergency care →; neither gives an ICU bed estimate directly. The serviceable market is a computed quantity: an annual per-bed platform fee multiplied by the bed count.
3. The deterministic CDS case
3.1 Regulatory argument
The FDA CDS guidance establishes that non-device CDS requires the clinician to be able to independently review the basis of the recommendation. Deterministic rule-based systems pass this test. Black-box AI systems do not.
The IMDRF SaMD Key Definitions and MDCG 2019-11 frame HealthOS ICU as software-in-device / accessory rather than standalone SaMD — reducing the regulatory burden of initial market entry.
3.2 Scientific argument
The closed truth-distillation cycle provides multicenter-ready reproducibility: identical compiled bundles deployed at multiple institutions produce identical outputs from identical inputs. This is a precondition for a well-controlled multicenter pivotal study.
3.3 Clinical argument
Phased rollout (capture only → internal review → recommendation only → supervised write → restricted write → full governance) provides the clinical governance trail that medical directors and ethics committees require. No institution is asked to accept autonomous device control before the evidence base is established.
4. Evidence requirements for stage transitions
Each transition in the six-stage validation protocol is governed by specific evidence requirements:
| Transition | Evidence type | Review body |
|---|---|---|
| Capture only → Internal review | Shadow recommendation quality metrics | Product team |
| Internal review → Recommendation only | Clinical reviewer consensus | Designated clinical reviewers |
| Recommendation only → Supervised write | Alert fatigue metrics + clinical authority approval | Clinical authority |
| Supervised write → Restricted write | Governed command-write incident rate + ethics committee review | Ethics committee |
| Restricted write → Full governance | Multicenter validation data | Independent clinical review |
The audit event log provides the data substrate for all stage-transition evidence reviews.
5. Research partnerships
HealthOS ICU is designed to support joint research in three modes:
Retrospective cohort studies. Historical episode data from the audit event log, de-identified and governed-exported, provides the substrate for retrospective analysis.
Prospective controlled studies. The platform can be deployed in capture-only or recommendation-only mode at a study site without affecting clinical practice; recommendations are compared against actual clinical decisions.
Interventional studies. Deployments at the supervised-write stage and beyond — where the platform’s recommendations directly influence clinical action — can be designed as controlled trials with the compiled bundle as the defined intervention.
All three study types benefit from the evidence-of-fact receipt as a publishable citation of the decision-support intervention.
Documents
Full Concept → · Regulatory and Geography → · Why Not AI →
Discuss a research partnership or data access for analysis. Contact us →