# Phase 17 — Public health surveillance, clinical escalation and responsible AI monitoring

## 1. Public health command centre
A date- and facility-filtered dashboard reports recorded:
- Disease surveillance notifications by condition.
- Suspected and confirmed case counts.
- Facility/condition notification counts.
- Maternal record registrations.
- Antenatal care visits.
- Child health registrations.
- Immunisation doses administered.
- Clinical alert counts.
- Open outbreak signals.

These are raw recorded service counts. The system does not have a population denominator, eligible-child cohort, vaccine schedule completeness or deduplicated epidemiological person-time model in this phase. Therefore it does not label these counts as immunisation coverage or disease incidence rates.

## 2. Outbreak early-warning signal
- Public-health administrators configure a condition label, threshold count, rolling window, classification scope and protocol reference.
- A scan counts matching surveillance notifications in the configured window.
- When the threshold is reached, the system records an outbreak signal for epidemiological review.
- Duplicate open signals for the same condition, facility and window are suppressed.
- Signals can be marked under investigation, verified, not an outbreak or closed.

**Safety limitation:** this is a configurable count threshold, not an epidemiological outbreak detector. It does not account for baseline incidence, seasonality, population size, age structure, geographic clustering, reporting delays or statistical confidence. A generated signal is not an outbreak declaration. Thresholds must be approved and interpreted by public-health professionals.

## 3. Case follow-up
Surveillance notifications can have follow-up records for:
- Case investigation.
- Contact tracing.
- Referral.
- Community alert.
- Follow-up.
- Other documented actions.

The records include outcome, next review time, user and facility. No SMS/email/public alert is sent by this workflow.

## 4. Clinical alert escalation
- Urgent/critical clinical alerts can be placed in an escalation queue.
- Queue entries specify escalation level, recipient role, message and facility.
- The designated role can acknowledge the escalation.
- Queueing is an internal record only; SMS, email, pager, push notification and guaranteed delivery are not implemented.

## 5. Responsible AI monitoring
- Added a monitoring page for model versions, manually entered performance metrics, prediction review status and AI incidents.
- Monitoring entries can document sensitivity, specificity, calibration, subgroup performance, drift or other approved metrics, along with method and notes.
- AI outputs remain advisory and subject to human review.

**Limitation:** this page records submitted measurements. It does not compute model performance, fairness, calibration, drift, confidence intervals or subgroup disparities automatically. It does not train, deploy or execute AI models.

## 6. Acceptance test source
- `PublicHealthSurveillanceAcceptanceTest`: threshold scan creates one review signal and suppresses duplicate open signal creation.
- `ClinicalAlertEscalationAcceptanceTest`: urgent/critical alert escalation is queued for a recipient role.

## 7. Operational and safety limitations
- No automatic disease diagnosis or outbreak declaration.
- No automated DHIS2/IDSR submission, e-reporting or external notification.
- No population denominators or immunisation coverage calculation.
- No geospatial maps or spatial cluster analysis.
- No external message delivery for clinical escalations.
- No automatic AI drift/fairness calculations.
- Public-health thresholds and clinical alert rules require approved protocols.
- Runtime migrations and PHPUnit tests have not been executed in this environment.
