AI Patient Journey Mapping

Jetstream2 AI-Pilot Fellowship

Dr. Greg Chism

U of A College of Information Science

Collaborator + Thank you

Dr. Jennifer De La Rosa

• Expert in Chronic Pain and Mental Health: Specializes in treatment quality, stigma, and health equity in chronic pain, substance use, and mental health.

• Leadership in Advocacy: Director of Strategy for the Comprehensive Center for Pain and Addiction and member of the U.S. Association for the Study of Pain Advocacy Committee and Leadership Academy.

• Accomplished Researcher: Recognized for impactful research, including an Editor’s Choice publication in PAIN.

• Workforce Development Leader: PI of PeerWORKS, training opioid-impacted individuals, and evaluator for AzCANN, providing cannabis education.

The Mental Health Disparity Challenge

The problem:

De La Rosa et al., 2024 Pain

• Chronic pain (CP) patients = significant barriers to mental health care

• Anxiety and depression symptoms often co-occur with CP

The Mental Health Disparity Challenge

Service Utilization Challenges:

Data Source: National Center for Health Statistics

Services are underutilized and fragmented, resulting in poorer outcomes.

The Mental Health Disparity Challenge

Significance

• Prevalence: High rates of co-occurrence between CP, anxiety, and depression.

  • Source: De La Rosa et al. (2024)

• Functional Impact: Disruptions in daily life and increased healthcare burdens.

  • Source: Kroenke (2023)

Call to Action

Leverage AI to analyze patient journeys, uncover disparities, and inform impactful interventions.

Call to Action: Similar Work

• Early Sepsis Prediction: Ensemble CNNs, LSTMs, and Transformer models to predict sepsis onset from time-series data, more accurate than traditional methods. Srivinas et al., 2024 JES

• Glucose Prediction: Hybrid Transformer-LSTM model predicts blood glucose levels with more precision than standard approaches. Bian et al., 2024 PLOS

• COVID-19 Impact Prediction: Assessment and comparison of LSTM networks and LSTM-CNN combinations to Transformer models with self-attention mechanisms for predicting COVID-19 impacts, with CNN-LSTM-based models outperforming others. Bharadi et al., 2023 IEEE Xplore

Call to Action: Novelty of This Project

• Targeting Care Gaps in Specific Comorbidities: Uniquely focuses on predicting care gaps in patients with comorbid anxiety, depression, and chronic pain, a critical area with documented unmet needs. ScienceDaily

• Ensemble Modeling of Patient Pathways: Applying an ensemble of LSTM and Transformer models to model patient pathways in this specific patient group is a novel approach, leveraging the strengths of both architectures to capture complex temporal patterns and interactions.

• Proactive Intervention Potential: By predicting care gaps, we aim to enable healthcare providers to intervene proactively, potentially improving patient outcomes and optimizing resource allocation in managing chronic conditions with psychological comorbidities.

Sorry for all of the text… 😮‍💨

Project Overview

Goal

Develop an AI-assisted tool to map and analyze patient journeys, focusing on individuals with CP, anxiety, and depression.

Scientific Contributions

1. Highlight mental health service utilization patterns and disparities.

2. Advance AI with a novel ensemble model of LSTMs and Transformers.

Impact

Provide actionable insights to improve accessibility and efficacy of mental health interventions.

Proposed AI Techniques

Sequence Models

• LSTMs: Capture short-term utilization patterns.

• Transformers: Extract long-term trends and dependencies.

  • ClinicalBERT: Fine-tuned for clinical text and irregular patient records.

Ensemble Model

• Combines strengths of LSTMs and Transformers.

• Identifies archetypal patient pathways for actionable insights.

Interpretability

• LIME: Ensures transparency and trust among stakeholders.

Target Data

• Focus: Critical care patients in ICU settings.

• Population: Data from the Beth Israel Deaconess Medical Center

• Limitations: Skewed toward critically ill patients and limited diversity in patient demographics.

• Strengths:

  • High granularity for ICU events.

  • Rich time-series data for vitals, labs, medications, and clinical notes.

• Focus: Broad population-level data across various healthcare settings.

• Population: Over 1 million diverse participants = comprehensive demographic and geographic representation

• Limitations: Data may have inconsistencies in granularity and standardization across health systems.

• Strengths:

  • Longitudinal data capturing routine care, chronic conditions, and diverse health behaviors.

  • Includes survey data (e.g., social determinants, health behaviors) that contextualize clinical data.

Technical Implementation

1. Data Preprocessing:

   • Clean, de-identify, and sequence MIMIC-III data for analysis.

2. Model Development:

   • Train LSTM and Transformer models (fine-tune ClinicalBERT).

3. Clustering:

   • Identify archetypes of service utilization pathways.

4. Stakeholder Collaboration:

   • Iterative model refinement with feedback from healthcare providers and policymakers.

Timeline

• Months 1–2: Data preparation, stakeholder consultations, computing setup.

• Months 3–4: Develop LSTM and Transformer models; initial analyses.

• Months 5–6: Apply ensemble models, analyze pathways, validate models.

• Month 7: Implement LIME, stakeholder review, and feedback integration.

• Months 8–9: Finalize analyses, prepare dissemination materials, stakeholder meetings.

Compute and Storage Needs

Compute

• 920 CPU hours (preprocessing and clustering).

• 770 GPU hours (model training and validation).

• Total: ~12,470 compute units.

Storage

• 1TB HIPAA-compliant storage (overestimation):

  • 500GB raw data.

  • 200GB preprocessed data.

  • 200GB model outputs.

  • 100GB results and documentation.

Expected Contributions

Domain Science

• Identify gaps and barriers in mental health service utilization.

• Guide policymakers and healthcare providers to improve care.

AI Advancements

• Develop a novel ensemble model for complex health data.

• Provide interpretable AI tools for practical applications.

Thank You 😊
Questions??