Project Overview:
This project involved developing a custom AI-driven system to monitor and predict International Normalized Ratio (INR) levels for a patient with an LVAD, addressing the critical need for precise anticoagulation management and proactive intervention. The system integrates daily logs of vitals and medications to provide faster and potentially more accurate INR predictions than traditional methods, ultimately aiming to maintain the patient’s INR within a therapeutic range.
Challenge:
The primary challenge stemmed from the inadequacies of the existing post-LVAD care protocols, which led to a severe thrombotic event. The client experienced dismissive responses to alarming LVAD data and a lack of proactive intervention from the medical team, resulting in delayed treatment and a life-threatening situation. This highlighted a critical need for a more responsive and data-driven approach to INR management to prevent future complications.
Solution:
To address these challenges, I developed a custom INR monitoring system leveraging AI and automation. The solution is designed to provide accurate and timely INR predictions, enabling proactive adjustments to medication, and seamless integration with other health management tools.
AI-Powered Interactive Assistant for Personalized Guidance
The core of the system is an AI model driven by OpenAI, accessible through an interactive chat interface. This allows for dynamic communication and personalized guidance. Users can ask specific questions, such as “If I eat spinach today, how should I adjust my Warfarin dose?” The AI considers the patient’s individual parameters and provides tailored recommendations, going beyond simple dosage adjustments.
This system is designed as a hematology agent within a broader AI care team, capable of taking on other tasks and integrating data from various labs to provide comprehensive hematology advice.
The system uses a Gravity Form interface within WordPress to input new INR values, timestamp, notes, and previous Warfarin dosage, which the AI then processes to generate dosing recommendations. This analysis allows the AI to predict future INR values, enabling proactive adjustments to Warfarin dosage. A critical innovation here is the introduction of a Warfarin Sensitivity Factor (WSF) to personalize dosing recommendations, combined with a detailed analysis of INR trends.
The WSF is calculated as:
WSF = (Change in INR) / (Change in Warfarin Dose)
By analyzing historical data, the system calculates the WSF for each dosage adjustment.
Example of WSF Calculation:
- Warfarin decreased from 4mg to 2mg, INR changed from 4.2 to 3.1.
- WSF = (-1.1) / (-2) = 0.55
The AI then uses this WSF to:
- Predict INR changes: Predicted Change in INR = (Change in Warfarin Dose) * WSF
- Recommend dose adjustments: Recommended Change in Warfarin Dose = (Target Change in INR) / WSF
Beyond the WSF, the AI incorporates a broader analysis of INR trends, including:
- Average INR and Warfarin Dosage: The system calculates the patient’s average INR and Warfarin dosage over a given period to identify overall control and dosage requirements.
- Warfarin Reduction Impact: The system analyzes instances where Warfarin was held or reduced to understand the patient’s response and the time course of INR changes.
- High INR and Dosage Adjustments: The system tracks instances of high INR values and the subsequent dosage adjustments to refine its understanding of the patient’s sensitivity and optimal correction strategies.
Real-time Feedback, Alerts, and Multi-Platform Access
The system provides immediate feedback upon form submission, displaying the AI’s Warfarin dosing recommendations, incorporating the WSF and trend analysis, and its rationale. This information is also automatically sent to the pharmacist for review.
If INR is 3.4, this is the feedback:
Daily Warfarin dose of 3.5 mg, yielding an INR of 3.4, which is within your target range of 2.5-3.5. Given that your INR is on the higher end of the target and considering your high sensitivity to Warfarin changes, the recommendation is to maintain the current dose of 3.5 mg daily without alteration. Projecting the INR trend, it is expected that the INR should stabilize around the upper end of the therapeutic range, assuming no significant dietary or medication changes that could affect Warfarin metabolism. The follow-up plan involves rechecking INR in one week unless you exhibit symptoms of bleeding or thrombosis, in which case immediate reassessment and possible dose adjustment will be necessary.”
If INR is 4.2, an alert SMS is triggered to the care team and this is the feedback:
Given the current INR of 4.2, which is notably higher than your target range of 2.5-3.5, it is advisable to temporarily hold your Warfarin dose for one day and recheck your INR within 24 hours. This adjustment is recommended to decrease the risk of potential bleeding complications associated with a higher INR. Closely monitor for any signs of bleeding and report any unusual symptoms to your healthcare provider promptly. Further modifications to your Warfarin dosage will be dependent on the follow-up INR result, aiming to maintain it within your therapeutic range to ensure stability and minimize risks.
In critical situations, where INR values fall outside the target range, the system triggers additional alerts:
- An alert email is sent to the relevant healthcare providers.
- Emergency alert SMS messages are sent to designated caregivers.
The system’s data, stored in Google Sheets, can be accessed and utilized across multiple platforms and by other AI assistants. For example, the patient can get a quick INR status update via Google Home. The Google Sheet can also be easily shared with healthcare providers and caregivers for collaborative monitoring. Furthermore, other AI assistants, such as a daily log assistant, can access and use the INR data and assistant to achieve their own goals, creating a connected ecosystem of health management agentic AI tools.
To ensure data privacy and HIPAA compliance, the system can be configured to store data locally using the wpDataTable plugin, providing an alternative to Google Sheets.
These real-time feedback, alert mechanisms, and multi-platform access, facilitated by WordPress automation and Google Suite integration, ensure timely intervention, proactive management, and seamless information sharing.
INR Log
Behind the Code: The AI’s Logic and Rule-Based System
The AI operates using a detailed protocol that mimics the reasoning of a specialized pharmacist. It considers factors like the patient’s LVAD type (Heartmate 3), medical history (including thromboembolism and bleeding risks), target INR range (2.5-3.5), and individual Warfarin response characteristics (sensitivity, clearance rate) as quantified by the WSF. To ensure consistent and reliable decision-making, the AI’s logic is structured around a series of “if-then” rules derived from the observed data patterns and key insights. This approach also helps to minimize reliance on more complex and potentially less predictable AI models, reducing computational cost.
In initial testing with other health AI, monthly costs reached approximately $10, prompting a shift towards more cost-effective and consistent rule-based systems. Furthermore, a local LLM (Ollama) is being explored as a potential alternative to cloud-based AI services, offering greater control and potentially lower costs. However, the current focus remains on refining the rule-based system for optimal performance.
[mwai_chatbot id=”default”]Examples of Derived Rules and Insights:
- Rule: If INR increases by >0.5 in 24 hours after a 0.5mg Warfarin increase, and the WSF is >1.0, reduce the subsequent dose by 0.25mg.
- Insight: “INR responds quickly to both stopping Warfarin (drops within a few days) and resuming at a high dose (increases rapidly). Even a 0.5 mg decrease (from 4 to 3.5 mg) significantly impacted INR, confirming sensitivity.”
- Rule: If INR is consistently within the 2.8-3.2 range, maintain the current Warfarin dose, even if outside the standard 2.5-3.5 range, to minimize fluctuations, and the WSF is <0.6.
- Insight: “The average INR was 3.13, which is within the target range (2.5–3.5) but tends to be on the higher end. The average Warfarin dose was 3.35 mg/day, though it fluctuated due to necessary adjustments.”
To provide accurate predictions, the AI uses an enhanced INR projection model that incorporates the WSF and trend analysis to estimate INR changes based on different Warfarin dosages. This model, combined with the refined rule-based system, allows for fine-tuned and personalized dose adjustments to maintain the patient’s INR within the therapeutic range.
Holistic Data Integration
Beyond just INR and Warfarin, the system integrates a comprehensive view of the client’s health status. It monitors timestamp, notes, heart rate, SpO2, blood pressure, other medications, and any complications. The data is organized and presented using interactive tables generated with the wpDataTable plugin within WordPress, providing a clear and accessible interface for tracking relevant information pulled from Google Sheets. This holistic approach provides a more complete picture of the client’s condition, enabling more informed decisions regarding INR management.
Targeted INR Range
The system is specifically programmed to maintain the client’s INR within a target range of 2.8-3.2, a narrower and potentially more therapeutic window than the general 2.5-3.5 range recommended by the doctors. This focus on a tighter range aims to minimize the risk of both thrombosis (clotting) and hemorrhage (bleeding), demonstrating a proactive approach to personalized medicine.
Results and Reflection:
The AI-powered INR monitoring system has demonstrated promising results in predicting INR values and enabling more proactive management of the client’s anticoagulation therapy. While still under refinement with ongoing data collection, the system has shown accuracy comparable to, and potentially exceeding, that of the medical team’s recommendations, with the added benefit of faster response times and critical alert functionality.
This project underscores the critical need for personalized medicine and the potential of AI to revolutionize patient care. It highlights the importance of empowering patients with tools to actively manage their health, especially in complex situations. The development of agentic AI tools, like the hematology agent described here, holds the promise of even more comprehensive and integrated healthcare solutions. In future iterations, I would aim to further refine the AI model, incorporate more data on dietary and lifestyle factors, and continue to prioritize cost-effectiveness and data privacy.
Technology:
- WordPress: The content management system (CMS) that provides the framework for the user interface and overall system functionality. I wield WordPress like a digital Swiss Army knife – it’s flexible, extensible, and allows for rapid development.
- Gravity Forms: A powerful WordPress plugin used to create the dynamic data input forms for INR values and notes. Gravity Forms isn’t just for simple contact forms; it’s a robust tool for capturing complex data.
- wpDataTable: An essential WordPress plugin that transforms raw data from Google Sheets into sortable, searchable, and visually appealing tables. It’s the key to presenting complex medical data in an understandable way to AI and for potential local storage to address HIPAA compliance.
- OpenAI: The driving force behind the AI model. By integrating with OpenAI’s API, I can leverage cutting-edge natural language processing to analyze patient data and generate intelligent insights.
- Google Sheets: Used as the backend data store for daily logs, INR logs, and medication logs. Google Sheets provides a flexible and collaborative way to manage the data that feeds into the system.
- Ollama (LLM): A local Large Language Model being used to enhance control, reduce costs, and improve consistency in AI decision-making.
