Building an AI-Powered Pain Management Platform: A Technical Case Study

The Pain Management Crisis in Numbers

Chronic pain affects 50 million Americans, according to the CDC. Of those, 19.6 million experience high-impact chronic pain that interferes with daily life. The annual economic burden of chronic pain in the U.S. exceeds $635 billion — more than cancer, heart disease, and diabetes combined.

Yet the tools clinicians use to manage pain remain largely paper-based questionnaires, subjective assessments, and one-size-fits-all treatment protocols. This case study documents how we built a digital pain management platform that combines FHIR-structured patient data, machine learning for pain trajectory prediction, multimodal treatment planning, and opioid risk scoring into a single clinical tool.

Our agentic AI solutions bring autonomous intelligence to healthcare operations.

Platform Architecture: FHIR-First Design

The platform was designed as a FHIR-native application from day one. Every piece of clinical data — pain assessments, treatment plans, medication orders, and outcomes — is represented as a FHIR R4 resource. This means the platform integrates with any FHIR-compliant EHR without custom mapping.

Core FHIR Resources

FHIR QuestionnaireResponse for Pain Assessment

{
  "resourceType": "QuestionnaireResponse",
  "questionnaire": "https://pain-platform.example.com/Questionnaire/brief-pain-inventory",
  "status": "completed",
  "subject": { "reference": "Patient/12345" },
  "encounter": { "reference": "Encounter/67890" },
  "authored": "2026-03-17T10:30:00Z",
  "item": [
    {
      "linkId": "pain-intensity-now",
      "text": "Rate your pain RIGHT NOW (0 = No pain, 10 = Worst pain imaginable)",
      "answer": [{ "valueInteger": 6 }]
    },
    {
      "linkId": "pain-worst-24h",
      "text": "Worst pain in the last 24 hours",
      "answer": [{ "valueInteger": 8 }]
    },
    {
      "linkId": "pain-location",
      "text": "Primary pain location",
      "answer": [{ "valueCoding": { "system": "http://snomed.info/sct", "code": "264186006", "display": "Lower back" } }]
    },
    {
      "linkId": "pain-interference-activity",
      "text": "How much has pain interfered with general activity?",
      "answer": [{ "valueInteger": 7 }]
    },
    {
      "linkId": "pain-interference-sleep",
      "text": "How much has pain interfered with sleep?",
      "answer": [{ "valueInteger": 5 }]
    }
  ]
}

Multimodal Pain Assessment: Beyond the 0-10 Scale

The traditional Numeric Rating Scale (NRS) — asking patients to rate their pain from 0 to 10 — captures a single dimension of a complex experience. Our platform collects data across four domains to build a comprehensive pain profile.

Domain 1: Patient-Reported Outcomes (PROs)

• Brief Pain Inventory (BPI) — Pain intensity and functional interference across seven domains
• PROMIS-29 — Physical function, anxiety, depression, fatigue, sleep disturbance, social participation, pain interference
• Pain Catastrophizing Scale (PCS) — Measures rumination, magnification, and helplessness related to pain
• Oswestry Disability Index (ODI) — For back pain patients: functional limitations in daily activities

Domain 2: Clinical Data from EHR

• Medication history — Current and historical analgesics, adjuvant medications, Morphine Milligram Equivalent (MME) calculations. Pulled from FHIR MedicationRequest.
• Diagnoses — ICD-10 pain-related diagnoses, comorbidities (depression, anxiety, substance use disorders)
• Procedures — Interventional procedures (injections, nerve blocks, spinal cord stimulation), physical therapy sessions
• Imaging — MRI, CT, and X-ray findings coded in SNOMED CT

Domain 3: Behavioral and Wearable Data

• Activity levels — Step count, active minutes, and mobility patterns from wearable devices via Apple HealthKit or Google Health Connect
• Sleep quality — Total sleep time, sleep efficiency, wake episodes from sleep trackers
• Heart rate variability (HRV) — A biomarker correlated with pain intensity and autonomic nervous system function

Domain 4: Psychosocial Factors

• PHQ-9 (depression) — Depression is present in 40-60% of chronic pain patients and directly impacts treatment outcomes
• GAD-7 (anxiety) — Anxiety amplifies pain perception and reduces treatment adherence
• Social determinants — Employment status, social support, housing stability affect pain management outcomes

ML-Powered Pain Trajectory Prediction

The platform uses machine learning to predict each patient's likely pain trajectory over the next 90 days. This allows clinicians to intervene early for patients whose pain is predicted to worsen, rather than waiting for deterioration.

Feature Engineering

The model ingests 78 features grouped into five categories:

Model Architecture

We use an ensemble approach combining gradient boosted trees (XGBoost) for tabular features with an LSTM (Long Short-Term Memory) network for temporal pain score sequences.

import xgboost as xgb
from tensorflow.keras.models import Model
from tensorflow.keras.layers import LSTM, Dense, Input, Concatenate

# Tabular features: demographics, medications, comorbidities
tabular_input = Input(shape=(52,), name="tabular_features")
tab_dense = Dense(128, activation="relu")(tabular_input)
tab_dense = Dense(64, activation="relu")(tab_dense)

# Temporal features: 90-day pain score sequence + daily activity
temporal_input = Input(shape=(90, 6), name="temporal_features")
lstm_out = LSTM(64, return_sequences=False)(temporal_input)

# Ensemble: combine both branches
combined = Concatenate()([tab_dense, lstm_out])
combined = Dense(64, activation="relu")(combined)
output = Dense(3, activation="softmax", name="trajectory")(combined)
# Classes: improving (0), stable (1), worsening (2)

model = Model(inputs=[tabular_input, temporal_input], outputs=output)
model.compile(optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"])

Model Performance

Opioid Risk Scoring: Data-Driven Prescribing Safety

With opioid-related overdose deaths exceeding 80,000 annually in the U.S. (CDC data), the platform includes a real-time opioid risk scoring engine that calculates risk at every prescribing encounter.

Risk Score Components

• Opioid Risk Tool (ORT) — Validated screening instrument assessing personal and family history of substance abuse, age, history of preadolescent sexual abuse, and psychological disease
• PDMP data — Prescription Drug Monitoring Program data showing concurrent opioid prescriptions, benzodiazepine co-prescriptions, and pharmacy shopping patterns
• Current MME — Morphine Milligram Equivalent dosage calculation. CDC guidelines flag doses above 50 MME for careful evaluation and above 90 MME for strong justification
• Concurrent medications — Benzodiazepines, muscle relaxants, gabapentinoids that increase overdose risk when combined with opioids
• Comorbidity flags — Active mental health diagnoses, sleep apnea (increased respiratory depression risk), renal or hepatic impairment affecting drug metabolism

FHIR RiskAssessment Resource

{
  "resourceType": "RiskAssessment",
  "status": "final",
  "subject": { "reference": "Patient/12345" },
  "encounter": { "reference": "Encounter/67890" },
  "occurrenceDateTime": "2026-03-17T10:30:00Z",
  "method": {
    "coding": [{ "system": "https://pain-platform.example.com", "code": "opioid-risk-v2", "display": "Composite Opioid Risk Score v2" }]
  },
  "prediction": [{
    "outcome": { "text": "Opioid-related adverse event within 12 months" },
    "probabilityDecimal": 0.72,
    "qualitativeRisk": { "coding": [{ "system": "http://terminology.hl7.org/CodeSystem/risk-probability", "code": "high", "display": "High Risk" }] }
  }],
  "note": [{ "text": "Risk factors: current MME 95, concurrent benzodiazepine, PHQ-9 score 18 (severe depression), prior opioid use disorder in remission" }]
}

Outcomes: What the Data Shows

After 12 months of deployment across three pain management clinics (2,400 patients), the platform demonstrated measurable improvements:

The most significant impact was the reduction in time to treatment adjustment. Continuous monitoring through the mobile app detected worsening pain trajectories an average of 11 days earlier than traditional follow-up visits, enabling proactive intervention.

Technical Implementation: Lessons Learned

Challenge 1: PRO Survey Fatigue

Patients asked to complete lengthy questionnaires stop responding within weeks. We implemented an adaptive questioning system that adjusts survey length based on the patient's trajectory. Stable patients get shorter surveys (3 questions). Patients with changing pain patterns get more detailed assessments.

Challenge 2: EHR Integration Variability

Different pain clinics use different EHR configurations. Some have structured pain assessment templates; others use free-text notes. Our EHR integration layer handles both cases by accepting FHIR QuestionnaireResponse resources from structured templates and using clinical NLP to extract pain data from unstructured notes when necessary.

Challenge 3: Model Bias in Pain Assessment

Pain assessment has documented racial and gender biases. Studies show that Black patients' pain is systematically underestimated by clinicians. Our ML model was trained on de-identified data from diverse populations and validated separately across demographic subgroups. We monitor fairness metrics (equalized odds across race and gender) in production and retrain quarterly.

Build a Pain Management Platform with Nirmitee

At Nirmitee, we specialize in building clinical decision support platforms that integrate with EHR systems through FHIR R4 APIs. Our team has experience with AI model development for clinical applications, patient-facing health dashboards, and scalable healthcare software architecture.

Need expert help with healthcare data integration? Explore our Healthcare Interoperability Solutions to see how we connect systems seamlessly. We also offer specialized Healthcare AI Solutions services. Talk to our team to get started.