Healthcare AI Innovation
RECOVER: Designing a Large Language Model-based Remote Patient Monitoring System for Postoperative Gastrointestinal Cancer Care
Authors: Ziqi Yang, Yuxuan Lu, Jennifer Bagdasarian, Vedant Das Swain, Ritu Agarwal, Collin Campbell, Waddah Al-Refaie, Jehan El-Bayoumi, Guodong Gao, Dakuo Wang, Bingsheng Yao, Nawar Shara
Affiliations: University of California, Irvine; Northeastern University; Johns Hopkins University; MedStar Health Research Institute; Creighton University; Georgetown University
Key Performance Metrics
RECOVER demonstrates significant improvements in patient monitoring accuracy, system usability, and safety, validated through extensive user studies and quantitative analysis.
0
Symptom Coverage (with CoT)
0
Clinical Staff SUS Score
0
Patient SUS Score
0
Assessment Errors / Hallucinations
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Empowering Postoperative GI Cancer Care with LLMs
RECOVER addresses the critical need for effective remote patient monitoring (RPM) in postoperative Gastrointestinal (GI) cancer care, a high-risk scenario due to unpredictable complications. Traditional RPM methods often suffer from survey fatigue, fragmented data, and heavy clinical workload, hindering timely interventions and increasing burnout.
Leveraging recent advances in Large Language Models (LLMs), RECOVER integrates a conversational agent for patients and an interactive dashboard for clinical staff. This system aims to provide clinically compliant, efficient, and responsible patient monitoring, bridging the gap between advanced AI capabilities and critical healthcare needs.
RECOVER System Architecture and Interfaces
RECOVER is designed with two primary user interfaces: a conversational user interface for patients (via Alexa Echo Dot) and a web-based dashboard for clinical staff. The system prompt for the conversational agent is structured with System Definition, Clinical Guidelines (based on 13 key symptoms), and Task Description, supporting empathetic, adaptive, and responsible patient interactions.
The backend includes an Information Extraction Module to categorize symptoms and extract Likert scale scores, and a Summarization Module to provide concise summaries of unstructured patient narratives. The provider dashboard offers a visualized patient list (sorted by severity), detailed patient information with key symptom visualizations (color-coded dots and meters), and a complete conversation log with highlights, enabling efficient review and action by clinical staff.
Validated Effectiveness and Usability
Pilot user studies with GI cancer patients and clinical staff confirmed RECOVER's high usability and effectiveness. Clinical staff reported an average SUS score of 93.75, praising the system for its user-friendliness and ability to prioritize critical cases. They found the LLM-powered features, including contextual insights and visualizations, highly supportive of their RPM work, allowing quick identification of urgent issues.
Patient participants rated the system with an average SUS score of 85, appreciating the "easy to use" interactions and the LLM's "top-notch" ability to understand human responses, comparing it favorably to traditional phone calls and questionnaires. The Chain-of-Thoughts prompting method significantly improved symptom coverage to 93.93%, and rigorous safety evaluations found zero cases of assessment errors or hallucinations.
Strategic Implications and Future Directions
RECOVER's design demonstrates that LLMs can effectively integrate domain-specific clinical guidelines and patient needs into RPM systems, providing robust support for clinical workflows. This can lead to a shift from reactive to sustained, preventative care, enhancing patient education and self-management.
Key considerations for future development include deeper integration with existing EHR/telehealth platforms, customization and personalization features based on patient recovery timelines, and comprehensive instructional materials. Emphasizing responsible AI, the system prioritizes accountability, clear capability boundaries, and patient privacy, with continuous human oversight crucial for high-stakes medical decisions.
Enterprise Process Flow: RECOVER Development Methodology
Participatory Design Study (5 Clinical Staff, 5 Patients)
→
Derive 6 Key Design Strategies
→
Design & Implement RECOVER System
→
Pilot User Studies (4 Clinical Staff, 5 Patients)
→
Identify Crucial Design Implications
0%
Reported Assessment Errors, Hallucinations, or Misunderstandings
A preliminary study showed RECOVER had no notable risks of assessment errors, misunderstandings, or hallucinations, highlighting its responsible AI design principles.
Comparison with and without Long-term Conversation History
| Exam-ple # | User Description of Symptoms | System with Long-term History | System without Long-term History |
|---|---|---|---|
| 1 | yes | It's good to hear from you today. I noticed from our recent conversations that you've been experiencing some pain. How are you feeling today - has your pain level changed since yesterday? | Great! Is any specific discomfort or symptom that you would like to talk about first? |
| 2 | pain in the lower abdomen | I'm sorry to hear you're experiencing pain in your lower abdomen. I remember you mentioned having pain in previous check-ins as well. Has this pain sharply increased or become unbearable since yesterday? | I am sorry to hear that you are experiencing pain in your lower abdomen. Does this pain increase sharply or become unbearable at times? |
| 3 | no I have not | Thank you for that information. I recall that you had some stomach pain about five days ago that was quite severe. How is your stomach feeling today? Are you experiencing any pain in that area now? | That's good to hear. Moving on, have you had any difficulty breathing lately? |
| 4 | no I have not | That's good to know. I remember that about five days ago, you mentioned having some stomach pain after your surgery. How is your stomach feeling now? Do you still have any pain that increases sharply or becomes unbearable? | Okay, that's good to know. Have you noticed any black, tar-like stools recently? |
| 5 | no | That's good to hear. I noticed from your records that earlier this month you had some difficulty tolerating food and drink. Have you been having any persistent constipation, nausea, or vomiting recently? | That's good to hear. Have you experienced any fever over 100 degrees or chills lately? |
| 6 | no | I understand. I notice you mentioned having difficulty swallowing yesterday as well, with a severity of 4 out of 10. Has there been any change in this throat discomfort since yesterday? | That's reassuring to hear. Are you noticing any wound drainage problems, such as redness, bleeding, pus, or an opening at the incision site? |
Case Study: User Experience & System Efficacy
Our pilot user studies involved both clinical staff and cancer patients, revealing RECOVER's substantial impact. Clinical staff highlighted the system's ability to help them prioritize critical cases and navigate patient data efficiently through LLM-powered summaries and color-coded visualizations. They noted the dashboard was "easy to use" and more straightforward than existing EMR platforms, boosting task efficiency.
Patients praised the conversational agent for its "top-notch" understanding and ease of use, preferring it over traditional communication methods. They experienced richer information exchange and appreciated the clinical integration that helped them understand their symptoms better. The studies confirm RECOVER's potential to significantly improve remote patient monitoring by fostering better patient-provider communication and streamlining clinical workflows, while emphasizing responsible AI practices to ensure safety and privacy.
Calculate Your Potential AI Impact
Estimate the efficiency gains and cost savings your enterprise could achieve by implementing an LLM-powered RPM system like RECOVER.
Estimated Annual Savings
Annual Hours Reclaimed
Your AI Implementation Roadmap
A phased approach to integrate RECOVER and other LLM-powered systems into your clinical workflows for maximum impact and minimal disruption.
Phase 1: Integrate with Existing EHR/Telehealth Systems
Seamlessly embed RECOVER's LLM conversation and dashboard modules into your current Electronic Health Record (EHR) and telehealth platforms for a unified experience.
Phase 2: Customization & Personalization Features
Develop adaptive interaction protocols, tailoring question depth and frequency based on individual patient recovery timelines, medical history, and specific needs.
Phase 3: Comprehensive Instructional Materials & Training Programs
Ensure smooth adoption by providing detailed guides and training for both patients and clinical staff, clarifying system capabilities and responsible AI usage.
Phase 4: Expand to Other Clinical Settings & Health Conditions
Generalize RECOVER's framework beyond postoperative GI cancer to address a broader range of RPM scenarios, enhancing its utility and scalability across healthcare.
Ready to Transform Your Patient Monitoring?
Discover how LLM-powered RPM solutions can enhance efficiency, improve patient outcomes, and reduce clinician burnout.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat