Agentic Workflow for Triggering Review of Systems Based on Treatment

Manual follow-up for medication side effects or treatment efficacy is a high-volume, low-complexity task that consumes nurse and MA time. This workflow automates the entire lifecycle: triggering questionnaires post-treatment, delivering them via patient portal or SMS, and summarizing responses for the chart. It directly saves 15-20 minutes of clinical staff time per monitored patient, reallocating human effort to higher-value care.

By systematically checking for side effects or lack of efficacy shortly after a treatment change, this workflow enables early intervention. It moves from a model where patients must self-identify and report problems to one where the practice proactively detects them. This can reduce emergency department visits or hospital readmissions related to adverse drug events, protecting patient safety and reducing total cost of care.

Structured, automated collection and clinical review of patient-reported data can qualify for Remote Therapeutic Monitoring (RTM) or Remote Physiologic Monitoring (RPM) billing codes under Medicare and many commercial plans. This workflow architectures the data collection, time spent, and documentation required to support compliant billing, turning a cost center (follow-up) into a revenue-generating service line.

Side effects are a primary driver of non-adherence. Proactive identification allows clinicians to adjust dosages, prescribe mitigations, or provide reassurance before the patient discontinues therapy. The workflow creates a continuous feedback loop between treatment and response, increasing the likelihood that patients stay on effective regimens, improving chronic disease management outcomes.

Every automated ROS response becomes a structured data point linked to a specific treatment and timeline. Aggregated across the patient population, this creates a powerful dataset for analyzing treatment effectiveness, identifying common side effect profiles, and supporting value-based care initiatives. It transforms episodic patient feedback into a longitudinal, queryable asset for clinical and operational decision-making.

The core orchestration logic—trigger (EHR event) → action (questionnaire delivery) → response (patient data) → triage (clinical review)—is a reusable pattern. Once built for one treatment (e.g., a new biologic), it can be extended to dozens of others across cardiology, oncology, rheumatology, and mental health. This creates a scalable platform for proactive care without proportional increases in administrative staff.

An autonomous agent integrated with the EHR (e.g., Epic, Cerner) that continuously scans for new or updated treatment plans (e.g., medication starts, dose changes). It identifies the specific treatment, maps it to a predefined library of potential side effects and efficacy timelines, and triggers the initiation of a time-delayed Review of Systems (ROS) questionnaire. This eliminates the manual, error-prone task of nurses or MAs tracking which patients need follow-up and when.

The core workflow logic, often built with frameworks like LangGraph or Temporal, that manages the patient journey. It schedules the ROS delivery (e.g., 7 days post-medication start) via the patient portal (or SMS/email), handles non-responses with gentle nudges, and adapts question branches based on initial patient answers. This engine integrates with patient engagement platforms (e.g., Luma Health, Phreesia) to execute communications while maintaining a unified audit trail.

A rules-based agent that analyzes completed ROS questionnaire responses. Using clinical logic (e.g., "reports severe headache + dizziness"), it scores the urgency, classifies responses into buckets like "Normal," "Monitor," or "Escalate," and routes them appropriately. Normal responses are logged in the EHR. Concerning responses are packaged with patient context and pushed into the clinical team's task queue in the EHR or a dedicated triage dashboard for timely review, preventing critical signals from being missed in inbox noise.

A system-to-system agent that handles bidirectional data flow. It structures patient-reported symptom data into a standardized format (e.g., FHIR resources) and writes a summary note or updates flowsheets in the patient's chart. It also listens for clinician actions (e.g., "patient called, advised to discontinue medication") and can trigger subsequent workflows, such as scheduling a telehealth visit. This closes the loop, ensuring the automated process feeds directly into the legal medical record and care continuity.

A critical control plane that logs every agent decision, questionnaire delivery, response, and routing action with timestamps and rationale. This creates a defensible audit trail for compliance (HIPAA, internal governance) and supports continuous improvement. It includes configuration dashboards for clinical leads to adjust escalation thresholds, update question libraries, and monitor workflow performance metrics (e.g., response rates, escalation volumes).

Not an agent, but a crucial component of the build. This refers to the phased deployment strategy, starting with a single treatment (e.g., a new hypertension medication) and one clinic. It involves configuring the initial treatment→symptom library, training staff on the triage dashboard, and establishing a feedback loop where clinical exceptions refine the agent logic. A successful pilot demonstrates ROI within 3-4 weeks, paving the way for scaling to dozens of treatments across the practice.

The workflow initiates when a new treatment plan (e.g., a medication order) is signed in the EHR (Epic, Cerner, Athena). An agent monitors the FHIR API or a dedicated event stream for specific RxNorm or SNOMED CT codes. Upon detection, it extracts the patient ID, medication details, and prescriber instructions to create a temporal follow-up schedule. This real-time integration eliminates manual chart review for new treatments.

A core LangGraph or temporal workflow engine manages the follow-up timeline. Based on the medication's known side-effect profile (from an internal knowledge base), it schedules a series of Review of Systems (ROS) questionnaires at critical intervals (e.g., 48 hours, 1 week, 1 month post-initiation). This engine handles state persistence, manages patient-specific delays (e.g., if a questionnaire is deferred), and triggers the next phase in the sequence.

Scheduled questionnaires are delivered via the patient portal (using its SMART on FHIR or REST API) or secure SMS/email. An LLM-powered conversational agent administers the ROS, using branching logic based on patient responses. Collected symptom data is structured into FHIR Observations and posted back to the EHR. For patients reporting severe symptoms (e.g., suicidal ideation, severe allergic reaction), the agent immediately triggers a HL7 ADT alert to the nursing triage queue.

Structured symptom data and severity scores are routed based on configurable rules. Normal or mild responses are logged to the chart. Moderate responses generate a task in the clinician's in-basket within the EHR. High-severity responses create a priority alert in the nurse triage module (e.g., Epic InBasket, TigerConnect) and can initiate an automated phone call to the on-call team. This ensures the right clinical resource addresses concerns at the right time, preventing adverse events.

All interactions, responses, and escalation outcomes are logged to a data warehouse (Snowflake, BigQuery). This fuels dashboards for population health managers to track medication tolerance rates and identify patterns. More critically, it creates a feedback loop: aggregate symptom data for a specific drug can refine the questionnaire timing and content in the knowledge base, making the system smarter and more precise over time.

Every automated action is recorded in an immutable audit log, linking patient ID, action type, data payload, and triggering logic. Access to configure medication-to-questionnaire mappings is role-based within an internal admin portal. Before go-live, the entire workflow is validated against a synthetic patient cohort in a sandbox EHR environment to ensure clinical safety and correct data flow. This governance layer is non-negotiable for regulated healthcare automation.