Government of Maharashtra | Project PRAGATI
AI CDSS Engine and Clinical Decision Trees
Decision-support engines for the five districts. Every AI output is an assistive recommendation that requires documented clinician or radiologist sign-off. No automated action on AI output is permitted.
Draft for discussion | October 13, 2026
Governance principle
Every AI output in this component is an Assistive Recommendation. It requires documented clinician or radiologist sign-off before any action. This aligns with the National Medical Commission (NMC) ethics guidance on AI and applicable Government of India AI principles (verify specific NMC reference before citing). Automated action on AI output alone is not permitted.
AI Chest X-ray Triage (TB, Pneumonia, Silicosis)
For high-burden areas: chest X-ray triage flags findings for radiologist review in priority order. The model outputs are bounding-box coordinates, a probability score, and a finding class. All outputs are assistive only.
X-ray triage flow
- 1
Patient presents at screening camp or facility
- 2
Chest X-ray acquired; DICOM image routed through PACS
- 3
AI model generates finding class, probability and bounding box
- 4
Radiologist priority queue updated (STAT / urgent / routine)
- 5
Radiologist reviews AI findings and completes report
- 6
Final report countersigned; AI finding stored as DICOM SR
Relevant national TB guidance should be consulted for algorithm configuration. No accuracy figures are stated because they are not sourced.
High-risk Pregnancy Stratification
Uses ANM / RCH / Anmol data alongside district hospital ultrasound records to produce risk scores for maternal and child health follow-up.
Pregnancy risk flow
- 1
ANM records antenatal visit and vitals into RCH / Anmol portal
- 2
Ultrasound data linked from district hospital
- 3
Risk engine computes stratification (low / medium / high)
- 4
High-risk cases flagged for specialist teleconsultation
- 5
CHW assigned for follow-up and counselling
Sepsis and ICU Early Warning Using MEWS and SOFA
Modified Early Warning Score (MEWS) based on vital signs from the HMIS, with SOFA score triggered for ICU patients. The MEWS grid below reproduces the published definition (Subbe et al., 2001). Scores are recomputed from structured data feeds and presented as an assistive recommendation.
| Parameter | Score 3 | Score 2 | Score 1 | Score 0 | Score 1 | Score 2 | Score 3 |
|---|---|---|---|---|---|---|---|
| HR (bpm) | <41 | 41-50 | 51-90 | 91-110 | 111-130 | >130 | - |
| RR (/min) | <9 | - | 9-11 | 12-20 | 21-24 | >24 | - |
| Temp (°C) | <35.1 | - | 35.1-36.0 | 36.1-38.0 | 38.1-39.0 | >39.0 | - |
| SBP (mmHg) | <91 | 91-100 | 101-110 | 111-249 | - | - | >249 |
| SpO2 (%) | <92 | 92-93 | 94-95 | >95 | - | - | - |
| AVPU | - | - | - | Alert | Verbal | Pain | Unresponsive |
Sepsis early warning flow
- 1
Vitals recorded at triage or nursing station and pushed to HMIS
- 2
MEWS score computed every hour for ward patients
- 3
For ICU patients, SOFA score computed daily
- 4
Score threshold triggers escalation protocol (nurse review, senior review, MET call)
- 5
Alert documented in EHR and routed to the treating clinician
- 6
Assistive recommendation displayed; no automated action
MEWS definition sourced from Subbe C.P., Kruger M., Gemmel L. (2001); NEWS2 definition from Royal College of Physicians (2017). SOFA scoring follows the Sepsis-3 consensus definitions (Singer et al., JAMA 2016). All scores are assistive only.
MJPJAY Claims FWA Anomaly Scoring
Flags suspicious claims against known fraud, waste and abuse (FWA) patterns for human adjudicator review. No automatic denial.
Claim anomaly flow
- 1
Pre-auth or claim submitted to NHA TMS / state portal
- 2
FWA engine scans against patterns (duplicate billing, high-frequency, outlier codes)
- 3
Anomaly score computed; claims above a threshold flagged
- 4
Flagged claims routed to human adjudicator at the State Health Assurance Society
- 5
Adjudicator reviews and makes final decision
Note:FWA anomaly scoring is advisory only. No claim is ever denied automatically. Pattern libraries are to be configured by the PMU.
Decision Flow Diagram: High-level
The diagram below shows the common decision flow for all five engines: