To develop AI-powered counselling module that goes beyond a chatbot and acts as an intelligent counselling assistant to the counsellor for TB patients and beneficiaries.

Tuberculosis counselling in the NTEP framework covers disease awareness, medication adherence, nutritional guidance, side-effect management, family screening, and government scheme linkage. The quality of this counselling is the single largest modifiable determinant of treatment.

PSMRI's STOP TB program deploys Counselling Officers across tribal geographies, many of whom are generalist health workers rather than trained TB counsellors. The NTEP mandates specific counselling protocols (CTB guidelines, DS-TB counselling modules, NPY guidance) but these exist as static documents not integrated into the point-of-care workflow.

Current Challenges

• Counselling quality varies significantly by individual worker experience and training recency.
• Counselling Officers must memorise or manually reference 60+ pages of CTB guidelines during sessions.
• No personalization: a severely malnourished 14-year-old DS-TB patient and a 52-year-old MDR-TB patient with diabetes receive identical generic counselling.
• Multi-language barrier: many Counselling Officers may not be the native speakers of the local tribal dialect.
• No real-time quality assurance hence supervisors cannot verify whether required counselling elements were delivered.
• Drug-food interactions, NPY linkage procedures, and TPT guidance for household contacts are frequently omitted.

Proposed Solution

An AI-powered Counselling Intelligence Engine embedded within the STOP TB application's Counselling Module. The engine operates as a silent clinical co-pilot not a chatbot, not conversational AI that reads the beneficiary's clinical profile in real-time and surfaces the most relevant, guideline-backed counselling content, structured recommendations, and audio-visual aids directly within the counselling form, without requiring any input from the Counsellor.

Principle

The AI does not replace the Counselling Officer it enables them. Every recommendation has to be traceable to a named guideline clause. The Counsellor retains full control: they can accept, modify, or override any recommendation. All AI outputs are explainable and source-attributed.

Objectives

• Deliver personalized, patient-profile-aware counselling guidance to every Counselling Officer at the point of care, regardless of experience level.
• Ensure 100% coverage of NTEP-mandated counselling elements for each patient category.
• Reduce the cognitive burden on field workers by surfacing the right content at the right moment in the session.
• Support multilingual counselling delivery through pre-produced audio-visual modules in multiple regional languages.
• Generate structured, auditable counselling summaries linked to each beneficiary record.

Key Requirements-

• The system should generate contextual counselling recommendations based on: Age, Gender, BMI/Nutritional status, TB disease type (DS-TB/MDR-TB/BPaL), Drug regimen, Comorbidities (HIV, DM, anaemia), Pregnancy status, Risk factors, Treatment stage etc.
• Counselling recommendations shall be sourced exclusively from approved NTEP/CTB guideline documents. No recommendation shall be AI-generated text — all outputs shall be direct excerpts with source attribution (document name, section number, page).
• The system shall support offline operation: all AI inference, guideline retrieval, and summary generation shall function without network connectivity.
• All AI models shall be explainable: every recommendation shall display its source clause and confidence rationale.
• The system shall be DPDP Act 2023 compliant: no beneficiary data transmitted to third-party AI services; all inference on-device.
• The system shall generate a structured Counselling Summary at session close, capturing: modules presented, guideline clauses surfaced, comprehension results, NPY linkage status, follow-up actions, and Counselling Officer override notes

Knowledge Base — Document Corpus

Offline and Low-Connectivity Considerations

• All AI inference runs entirely on-device in an offline setting.
• Knowledge base vector index (~18MB) and all NTEP guideline text are stored locally.
• Model updates (new CTB guidelines, revised drug-food interactions) are pushed as differential update packages on next sync — no full reinstall required.
• Counselling Summaries and all AI outputs are stored in the AMRIT local servers and synced to central on connectivity restoration.

AI-Powered Counsellor Performance Feedback

This module extends the same Counselling Intelligence Engine described above to the Counsellor's own delivery of the session, rather than to the beneficiary. The AI analyses how the counselling was delivered, not what the beneficiary said or disclosed, and operates under the same explainability and human-control principles as the recommendation engine above.

How It Works

During the session, audio is captured only after the Counsellor has started a feedback-enabled session and the beneficiary has been informed and has given consent. On-device speaker diarisation separates the Counsellor's voice from the beneficiary's voice. Only the Counsellor's speech is transcribed and analysed against the structured counselling checklist that also drives the recommendation engine; the beneficiary's speech is used only transiently, to detect conversational markers such as a question being raised and a response being given, and is not retained or analysed for content. Analysis can run in real time, flagging protocol gaps the Counsellor can still act on before the session ends, or as a post-session process, depending on device capability and programme preference.

What the AI Analyses

• Completeness: which mandatory counselling elements for that patient's profile were actually covered versus skipped.
• Protocol adherence: whether the content and sequence of what was said maps to the recommended CTB/NTEP counselling script for that case type.
• Clarity of communication: sentence length, jargon density, pace, and whether explanations were simplified appropriately for the beneficiary.
• Empathy and tone: vocal markers such as speaking pace, interruptions, and pauses left for questions.
• Use of simple/local language: whether the session was delivered in the beneficiary's regional language or dialect and avoided untranslated clinical terms.
• Key topics covered or missed, mapped against the checklist (nutrition, drug-food interaction, NPY linkage, TPT for household contacts, etc.).
• Whether beneficiary concerns were addressed, inferred from conversational turn-taking (a question followed by a substantive answer) without transcribing or storing what the beneficiary actually said.

Post-Session Feedback

At the close of each session, the Counsellor receives a short, private feedback summary on their own device, covering:

• What was done well
• Areas of improvement
• Topics or counselling points missed
• Communication and behavioural observations
• Specific, practical suggestions for the next session

As with the recommendation engine, feedback is written in plain, constructive language and references the specific guideline clause or checklist item it relates to, so it remains explainable and non-arbitrary rather than a generic score.

Continuous Improvement

Individual session feedback accumulates into a private trend view for each Counsellor: consistency of protocol adherence over time, recurring gaps, and topics repeatedly missed. This lets a Counsellor track their own progression and gives programme teams a basis for targeted refresher training and mentoring, rather than one-off correction, so counselling quality improves progressively and consistently across the workforce.

Non-Judgemental, Coaching-First Approach

This module is positioned as a private coaching layer for the Counsellor, not a surveillance or a performance-rating tool. Feedback is delivered to the Counsellor first and, by default, is visible only to them; sharing beyond that, for example with a supervisor for mentoring, should be an opt-in or programme-configured setting rather than an automatic default. No individual session recording, transcript, or score should be usable on its own as a disciplinary record.

Privacy, Consent, and Data Protection Safeguards

• Explicit consent is obtained from the beneficiary before any audio capture begins, and the beneficiary may decline without it affecting their counselling or care.
• All speech processing, diarisation, transcription, and analysis, runs on-device, consistent with the existing offline, no-third-party-transmission design of the Counselling Intelligence Engine.
• Beneficiary speech is not stored or transcribed; only the Counsellor's speech and the resulting protocol-adherence analytics are retained.
• Any raw audio retained for quality assurance is time-limited, encrypted at rest, and auto-purged after a defined retention window.
• Access to individual feedback is restricted to the Counsellor themselves; supervisor or programme access is role-based, limited to aggregated or explicitly shared data, and logged.
• The module remains DPDP Act 2023 compliant, in line with the rest of the Counselling Module.

Supervisor Dashboard and Aggregated Analytics

Supervisors and programme managers access only aggregated, de-identified counselling-quality metrics, for example protocol adherence rates by facility, region, or Counselling Officer cohort, common topics missed, or language and dialect gaps, without access to raw recordings, transcripts, or individual conversation content. These aggregated insights feed directly into training prioritisation and quality-assurance reporting, complementing the auditable counselling summaries already generated by the Counselling Intelligence Engine.

Expected Benefits and Impact

• Guideline-backed support without memorization; 40–60% reduction in time per session.
• Personalized, culturally appropriate counselling in local language; higher comprehension.
• Auditable counselling records; quality gap identification; training prioritization.
• Real-time counselling quality data at national scale; evidence for scale-up decisions.