Predictive attrition risk
Phase 2 Roadmap — Mock
This is a non-functional preview. Live attrition modelling is not yet implemented — the numbers below are illustrative.
Predict who will leave — before they do
A federated model trained on tenure, supervision frequency, geography, and engagement signals. Surfaces at-risk workers 60–90 days ahead so the SPHCDA can intervene — supervision visit, posting review, or top-up disbursement — before the post becomes vacant.
Lafia
HIGH8 of 21 CHEWs at risk
Model confidence: 84%
·Longer-than-average post tenure
·Distance from supervisor hub
·Wage parity gap
Karu
MEDIUM3 of 17 RMs at risk
Model confidence: 71%
·Recent FCT private-sector hiring spike
·Limited continuing education
Akwanga
LOW1 of 12 nurses at risk
Model confidence: 65%
·Stable team, near training hub
Workers flagged for proactive outreach
Ms A.M.CHEW· Akurba PHC, Lafia
12 months in post — past common exit window
0 supervision visits in last 90 days
Lives 35 km from facility (no staff quarters)
Risk score
78
Mr K.O.RM· Mararaba PHC, Karu
Applied for FCT transfer twice in 6 months
Recently completed midwifery upgrade
Below-average BHCPF top-up arrears
Risk score
62
Mrs N.A.JCHEW· Wakama PHC, Doma
Currently on maternity leave, second child
Spouse relocated to Abuja
No childcare support available locally
Risk score
55
Mr T.D.RN· Toto General Hospital
9 months without promotion despite eligibility
Reported burnout in last engagement survey
Risk score
48
Federated learning
Each SPHCDA trains a local model on its own FHIR registry. Only model weights — never PII — are shared back to NPHCDA for aggregation.
60–90 day horizon
The window where targeted retention spend actually pays off. Earlier is noisy; later is too late.
Intervention catalogue
Each flag links to a specific action: supervision visit, hardship top-up, training pull-up, posting review.
Production architecture (planned)
FHIR PractitionerRole + tenure history → feature pipeline (Kafka) → per-state PyTorch trainer → DP-aggregated weights → NPHCDA inference Worker → write back as RiskAssessment resources tagged to each Practitioner. No personally identifying data leaves the state PHCDA boundary.