Population Health
Risk stratification, prediction, and evidence-based outreach using population-level clinical and sociodemographic data.
Developer Console Dashboard
The Developer Console includes a population health module that brings surveillance, forecasting, intervention planning, and campaign tracking into a single workflow. The module is organized around six pages that mirror the public health decision cycle: identify risk, project its trajectory, simulate responses, deploy campaigns, and report outcomes.
Overview
The Overview page is the default landing view. It displays territory-wide risk indicators, a list of active campaigns with their current status, flagged districts that need attention, and a watchlist of populations the system is tracking. Operators use this page to orient before drilling into specific areas.
Surveillance
The Surveillance page shows current disease indicator rates on a geographic choropleth map with district-level granularity. It includes anomaly detection and ranking, spatial cluster identification, multivariate alert correlation, and a time scrubber for replaying historical indicator states. A playbook graph recommends response sequences for detected anomalies.
Forecast
The Forecast page projects disease incidence forward using trained models. A fan chart shows the median forecast with confidence intervals, and operators can animate the projection month-by-month to see how the model's fitted trajectory extends from historical observations into the forecast horizon. A cost-of-inaction card quantifies projected additional cases if the current trajectory continues without intervention, linking directly to the intervention simulator. A stratified-fits heatmap shows model discrimination across cohort and region combinations.
Interventions
The Interventions page is where operators simulate and compare potential health interventions before deployment. Operators select from a catalog of evidence-rated interventions, adjust deployment intensity, and see projected cases prevented, cost, and cost-per-case update in real time. A leaderboard ranks interventions by projected impact. Operators can pin up to three interventions to a comparison tray for side-by-side evaluation across cases prevented, percentage reduction, total cost, and cost efficiency. A geographic playback view shows how district-level rates would evolve over a 36-month horizon under the selected intervention. An evidence feedback panel shows how results from deployed campaigns update the evidence base for each intervention.
Campaigns
The Campaigns page tracks interventions that have been deployed. Each campaign displays a trajectory chart comparing the no-action counterfactual (what would have happened), the original forecast, and observed reality since deployment. The shaded area between observed outcomes and the no-action line represents estimated cases averted. Campaign cards show deployment metadata including district, enrolled cohort size, duration, and equity status.
Policy Memo
The Policy Memo page generates structured district-level briefs summarizing risk status, active interventions, projected outcomes, and equity considerations. These briefs are designed for district-level decision makers who need actionable summaries without navigating the full dashboard. and Risk Management
This guide covers how Amigo supports population health workflows: ingesting clinical and sociodemographic data at scale, stratifying patients by risk, predicting outcomes across clinical indications, and delivering evidence-based recommendations through proactive agent-driven outreach.
The Challenge
Health systems sitting on large volumes of clinical and sociodemographic data face a common problem: the data exists, but it does not drive action. Risk models run in analytics platforms that are disconnected from patient communication. Stratification scores sit in dashboards that care coordinators check manually. Evidence-based guidelines live in PDFs and clinical protocols that staff apply inconsistently.
The gap is not data or knowledge. It is execution: turning a risk score into a patient conversation, turning a guideline into a recommendation delivered at the right moment, doing this across thousands of patients without scaling staff linearly.
Amigo connects the data infrastructure (ingestion, unification, risk computation) directly to agent-driven conversations (voice, text, surfaces) in a single platform.
Data Foundation
Population health starts with data. The platform needs a unified view of each patient that spans clinical records, sociodemographic context, and interaction history.
Clinical Data Ingestion
The connector runner pulls clinical data from EHR systems, FHIR stores, and other source systems. For population health, the relevant data includes:
Diagnoses and problem lists: active conditions, comorbidities, historical diagnoses
Lab results and vitals: HbA1c trends, blood pressure readings, lipid panels, BMI
Medications: current prescriptions, adherence patterns, refill history
Encounter history: visit frequency, no-show patterns, ED utilization, hospitalizations
Care plans and orders: open referrals, pending follow-ups, screening gaps
The connector runner supports multiple connector types: direct FHIR integration, REST API polling, file drops for batch imports, and webhooks for real-time updates. Each data source feeds into the world model as confidence-scored events. An EHR diagnosis arrives at 1.0 confidence. A patient-reported symptom from a voice call arrives at 0.5. The world model's projection function computes each patient's current state from all events, resolving conflicts by confidence rather than timestamp.
Sociodemographic Data
Clinical data alone misses critical risk factors. The same connectors that pull clinical data can ingest sociodemographic indicators:
Demographics: age, sex, race/ethnicity, preferred language
Social determinants: housing stability, food security, transportation access, employment status
Geographic context: ZIP code, urban/rural classification, distance to care facilities
Insurance and financial: coverage type, plan limitations, prior authorization patterns
Behavioral factors: smoking status, substance use, physical activity levels
Social determinant data often arrives from multiple sources with varying reliability. A patient's housing status might come from a social work assessment (high confidence), a screening questionnaire during a call (medium confidence), or an agent inference from conversation context (lower confidence). The world model's confidence hierarchy picks the most reliable source without discarding less certain signals.
Unified Patient View
The world model unifies all of this into a single event-sourced record per patient. Clinical, sociodemographic, and interaction data coexist in the same entity model. This is the foundation that makes downstream risk computation and agent-driven outreach possible. The agent does not need to query five systems to understand a patient's full picture.
Risk Stratification
Once data is unified, the platform computes risk scores that stratify the population.
Platform Functions for Risk Scoring
Platform functions are where risk models run. The three function types map to different stratification needs:
SQL Table
Cohort queries that identify patients meeting clinical criteria (e.g., all diabetics with HbA1c > 9 and no visit in 90 days)
Python UDF
Custom scoring algorithms that combine clinical indicators, sociodemographic risk factors, and utilization patterns into composite risk scores
AI Built-in
Classification functions that categorize patients by acuity or urgency from unstructured clinical notes
A Python UDF for risk scoring might weight lab values, comorbidity count, social determinant flags, and care gap indicators into a composite score per patient. This function runs on the platform's compute layer with no external model hosting or separate analytics pipeline required.
The built-in fn_risk_score function computes composite risk from multiple signals. For condition-specific models, teams register custom functions that encode their clinical logic and evidence-based scoring criteria.
Condition-Specific Stratification
Each clinical indication can have its own risk model registered as a named function. A diabetes risk stratification function weights HbA1c trends, medication adherence, retinal screening gaps, and social determinants differently than a cardiovascular risk function that weights lipid panels, blood pressure trends, smoking status, and family history.
The catalog discovery mechanism means new risk functions become available to agents automatically once registered. No deployment pipeline or agent reconfiguration required.
Population-Level Queries
The open query tool lets agents construct ad-hoc SQL at runtime to answer population-level questions: "How many patients in this panel have uncontrolled diabetes and a transportation barrier?" or "Which patients scored high-risk but have not been contacted in the last 30 days?" This enables dynamic cohort identification during operational workflows.
Epidemiology Tables
Risk stratification is only trustworthy if it is explainable. Why does one cohort score high? Why does another drift? The platform produces a layered set of population-health tables - built from the world model's clinical projections - that surface the underlying signal instead of hiding it behind a black-box score:
Prevalence with small-sample confidence intervals. Active-condition rates come with Wilson 95% intervals that stay sensible even for tiny cohorts, so a prevalence curve on a small panel does not silently mislead.
Cohort drivers via Standardized Prevalence Ratio (SPR). For every condition × cohort pair, the system publishes whether the cohort over- or under-indexes for that condition, with Byar 95% intervals. This is the interpretable "why this cohort ranks high" signal - there is no feature-importance guessing, just a standard epidemiology measure with a CI around it.
Per-patient Charlson Comorbidity Index. Quan (2005) ICD-10 mappings and Charlson (1994) age-adjustment produce a feature-ready burden score for every patient. Downstream risk models - whether MLflow pipelines, SQL rules, or off-the-shelf AutoML - consume this as one of their features. The burden computation itself is deterministic and auditable.
Co-occurrence pattern mining. Jaccard similarity, lift, pointwise mutual information, and chi-square statistics over condition pairs drive comorbidity pattern discovery. Every pair carries a small-cell suppression flag so UIs can filter out unstable pairs automatically.
Incident trends by month and cohort. Incident case counts bucketed by month × code × cohort feed outbreak detection and seasonality analyses.
All measures are methodologically cited inline (Wilson 1927, Breslow & Day 1987, Charlson et al. 1987/1994, Quan et al. 2005) - this is the same statistical toolkit NCI/SEER and CDC use for public surveillance reporting. The explainability bar is "can a clinician or auditor read the formula," not "can a data scientist reverse-engineer the model."
Risk Prediction
Stratification tells you where patients are today. Prediction tells you where they are heading.
Memory-Driven Predictive Signals
The platform's memory system extracts structured facts across six built-in dimensions for every patient:
Clinical
Diagnoses, medications, allergies, symptoms
Behavioral
Communication preferences, emotional baseline, response patterns
Operational
Scheduling patterns, no-show risk, preferred appointment times
Social
Caregivers, living situation, emergency contacts
Engagement
Channel preferences, response rates, last contact
Risk
Acuity level, adherence risk, care gap score
These dimensions are workspace-configurable. Teams can adjust weights, add custom dimensions for specific clinical indications, or use LLM-powered semantic extraction for novel risk factors.
Over time, per-patient trajectories across these dimensions reveal predictive patterns:
L1 (Extracted Memories) captures facts and preferences from each interaction: medication concerns, reported barriers, symptom changes
L2 (Episodic Models) synthesizes per-patient patterns across encounters: declining engagement, worsening symptom reports, increasing no-show frequency
L3 (Global Model) aggregates population-level patterns: which dimensional profiles predict deterioration, which intervention sequences correlate with improvement
Dimension completeness scores per patient show where the model has strong signal and where gaps remain. A patient with high clinical dimension coverage but sparse social data may need a social determinant screening before the risk model can produce a reliable score.
The analytical capabilities built on this memory data enable predictive population intelligence:
Dimensional analysis identifies which patient attributes most strongly predict outcomes across the population - for example, that emotional well-being dimensions predict medication adherence better than clinical knowledge
Cohort performance analysis compares outcome trajectories across patient segments to identify which groups are diverging toward poor outcomes
Causation lineage analysis traces how early interventions cascade into long-term results, quantifying which touchpoints matter most
Continuous Monitoring
The outbound system's continuous monitoring pattern evaluates patient state on a recurring basis. When a patient's risk trajectory changes (adherence drops, care gaps accumulate, engagement declines), the platform creates outbound tasks automatically. This turns prediction into action without manual chart review.
Evidence-Based Recommendations
Risk scores and predictions are only valuable if they drive the right clinical action. The platform delivers evidence-based recommendations through agent conversations.
Knowledge System for Clinical Protocols
The knowledge system embeds clinical protocols and evidence-based guidelines directly into the agent's reasoning through dynamic behaviors. Rather than storing guidelines as static reference material, the system activates the relevant clinical framework at the right moment in a conversation.
For each clinical indication, dynamic behaviors encode:
Screening protocols: when and how to assess the patient against evidence-based criteria
Intervention pathways: which recommendations to deliver based on risk level and patient context
Escalation boundaries: when the agent must defer to a clinician rather than deliver a recommendation
Compliance constraints: regulatory requirements that govern what the agent can and cannot say
The high-stakes protocol framing pattern ensures the agent applies clinical frameworks automatically when the conversation enters high-risk territory, without explicit instruction.
Personalized Delivery
Evidence-based does not mean one-size-fits-all. The agent adjusts how it delivers recommendations based on the patient's full context:
Health literacy: The agent adjusts language complexity. A recommendation about HbA1c management sounds different for a nurse than for a patient with limited health literacy.
Barriers and preferences: If the world model shows transportation barriers, the agent factors that into care recommendations (telehealth options, pharmacy delivery, closer facilities).
Prior interactions: The memory system tracks what the patient has tried before, what worked, and what they resisted. Recommendations build on this history rather than repeating failed suggestions.
Cultural context: Language preferences, communication style, and cultural factors from the patient's profile shape how the agent frames recommendations.
Progressive Data Collection
When the agent identifies gaps in the patient's record during a risk assessment conversation, it can generate a surface: a data collection form delivered via SMS, WhatsApp, email, or web. Surfaces capture structured data (screening questionnaires, social determinant assessments, symptom checklists) that flows back to the world model and refines the patient's risk profile.
Data collection is progressive rather than all-at-once. A first touchpoint might collect two fields (name and date of birth). The next surface arrives prefilled from the first response plus any data already in the world model, so the patient only fills in what is new. By the time a full intake form is presented, most fields are already populated. Fields auto-save as the patient fills them in, so partial completions still provide usable data. Saved progress expires after a configurable period, giving patients a reason to finish before their data is cleared.
Proactive Outreach at Scale
The platform does not wait for patients to call. The outbound system uses risk stratification to drive engagement across multiple channels.
Multi-Channel Engagement
Outreach spans voice calls, SMS, WhatsApp, iMessage, RCS, email, smart forms, portal notifications, voicemail drops, and direct mail. Each channel serves a specific purpose. SMS handles quick prompts and micro-asks. Email provides space for education and benefit breakdowns. Phone calls handle objections in real time and close conversions. Direct mail reaches patients who do not engage with digital channels. The platform selects the right channel for each patient and each phase of engagement based on observed behavior, not a static sequence.
Adaptive Outreach
The system watches what patients do, not just whether they respond. Opening an email but not clicking through is a different signal than ignoring all digital contact entirely. Starting a form but abandoning halfway is different from never clicking the link. Each of these signals triggers a different response:
A patient who opens emails but does not click gets a different message framing on the next touch, not a repeat
A patient who starts a form but abandons gets a resume prompt ("you are almost done") through a different channel
A patient with zero digital engagement gets a physical postcard with a QR code
A patient who answers a call but says "not now" gets a callback scheduled at their preferred time
This behavioral branching means the channel and message adapt to each patient's engagement pattern. The platform does not repeat what did not work.
Campaign-Driven Outreach
Orchestrated campaigns target patient populations based on risk scores and clinical criteria. Examples across clinical indications:
Contact all high-risk diabetic patients due for HbA1c testing
Reach patients with cardiovascular risk factors who missed statin follow-up
Schedule screening calls for patients flagged by a behavioral health risk model
Follow up with post-discharge patients whose readmission risk score exceeds a threshold
Campaigns have configurable priority and pacing. A high-risk outreach campaign runs at higher priority than a routine wellness check. The system respects business hours, contact preferences, and retry logic per patient.
Compliance Gates
Every outbound message passes through a series of compliance checks before delivery. All gates are fail-closed: if any check fails, the message is skipped, not sent.
Consent verification: opt-in confirmed per channel, with immediate permanent opt-out on request
Quiet hours: no contact outside permitted hours in the patient's local timezone
Cooldown periods: minimum intervals between outreach per patient to prevent over-contact
Fatigue gating: maximum pending surfaces per patient, with automatic suppression when completion rates drop below threshold
Human review: operator approval queue for outreach that meets review criteria
Outreach Optimization
The platform tracks per-patient surface completion rates, preferred channels, and pending form counts. Over time, this builds a per-patient engagement profile:
Channel effectiveness: which channels produce the best response for each patient
Timing patterns: when the patient is most likely to engage
Completion behavior: how far through forms the patient typically gets
These signals feed back into campaign prioritization so outreach effort is concentrated where it is most likely to produce action. The system learns what works for each patient specifically, not for patients in general.
Full-Loop Workflow
The end-to-end flow for a clinical indication:
Connector runner loads clinical and sociodemographic data from EHR and other sources
World model unifies patient records with confidence-scored events
Platform functions compute risk scores per patient per clinical indication
Outbound system creates tasks for patients exceeding risk thresholds
Agent calls or texts the patient with full context pre-loaded
Dynamic behaviors activate the relevant clinical protocol and evidence-based recommendations
Agent delivers personalized guidance, collects new data, schedules follow-up if needed
World model records interaction outcomes as events, updating the patient's risk trajectory
Connector runner syncs verified outcomes back to the EHR
Each step is auditable. Every risk score computation, outbound attempt, recommendation delivered, and patient response is recorded as a world model event with full provenance.
Verification Across Clinical Indications
Each clinical indication requires its own verification before reaching patients. The platform's testing framework supports this through:
Condition-specific simulation personas: synthetic patients representing the clinical and sociodemographic diversity of each indication's target population
Clinical scenarios that test risk-appropriate responses, where the agent must handle a low-risk patient differently than a high-risk one
Safety gates: 100% escalation accuracy, 100% scope-of-practice adherence, 99.5%+ medical information accuracy per indication
Drift detection that monitors whether agent performance changes as the patient population evolves
For detailed guidance on building clinical verification suites, see Clinical Verification.
Multiple Clinical Indications
The platform supports multiple clinical indications simultaneously through its compositional architecture:
Separate context graphs per indication define the conversation flow and clinical boundaries for each condition
Indication-specific platform functions encode the risk model and scoring criteria for each condition
Dedicated dynamic behaviors carry the evidence-based protocols for each indication
Shared data infrastructure: the world model, connectors, and outbound system are shared across indications, so a single patient can be stratified across multiple conditions simultaneously
Adding a new clinical indication does not require rebuilding the platform. It requires defining the risk model (platform function), the clinical protocol (dynamic behaviors and knowledge), the conversation flow (context graph), and the verification suite (simulations and metrics). The data infrastructure, outbound system, and agent engine are already in place.
See also
World Model for how clinical and sociodemographic data is unified
Platform Functions for risk scoring and population queries
Analytical Capabilities for cohort-level predictive intelligence
Knowledge for embedding clinical protocols in agent reasoning
Outbound System for proactive patient outreach
Clinical Verification for testing clinical safety per indication
Last updated
Was this helpful?