AI-Enhanced Healthcare Process Mining: New Framework Shows Promise for Clinical Pathway Analysis

In a breakthrough that could transform how healthcare professionals analyze complex patient data, researchers have developed a new framework that combines process mining with artificial intelligence to make sense of complicated healthcare workflows. The system, called HealthProcessAI, uses large language models to interpret complex healthcare processes and translate them into understandable reports for clinicians and researchers.

Bridging the Gap Between Data and Clinical Insights

Healthcare systems worldwide generate massive amounts of electronic data through patient records, monitoring devices, and clinical information systems. Despite this wealth of information, healthcare professionals often struggle to extract meaningful insights from these complex datasets.

"Process mining has shown tremendous potential for understanding healthcare workflows, but its adoption faces significant barriers," explains Eduardo Illueca Fernandez, lead author of the research published on arXiv. "Many clinicians lack the technical expertise to use existing tools, and interpreting the results requires specialized knowledge that most healthcare professionals don't have."

The new framework addresses these challenges by creating an educational wrapper around established process mining libraries and integrating AI language models to automatically interpret the results. This approach could make powerful analytical techniques accessible to a much wider audience of healthcare professionals.

"What makes HealthProcessAI unique is its focus on accessibility and education," says Kaile Chen, co-author of the study. "We've designed it not just as an analytical tool but as a learning platform that guides users through the process while maintaining scientific rigor."

The researchers tested their framework using data related to sepsis progression and kidney disease, demonstrating how it can identify critical patterns in patient care pathways that might otherwise remain hidden in complex datasets.

How the System Works

HealthProcessAI follows a six-step modular pipeline that transforms raw healthcare data into meaningful insights:

1. Data Loading and Preparation: The system prepares healthcare data according to clinical standards, ensuring compatibility with diverse information systems.

2. Process Mining Analysis: Using established algorithms, the framework analyzes event logs to discover patterns and pathways in patient care.

3. LLM Integration: Five different large language models interpret the technical outputs, translating complex analyses into understandable reports.

4. Advanced Analytics: The system applies research-grade methodologies to identify bottlenecks, predict outcomes, and check conformance with clinical guidelines.

5. Report Orchestration: Multiple AI models work together to create comprehensive reports that capture diverse perspectives and insights.

6. Validation Framework: The system includes built-in validation tools to ensure accuracy and reliability.

"Each component serves a specific purpose in making process mining more accessible and useful for healthcare applications," notes Fernando Seoane, who contributed to the research. "The modular design allows users to customize the analysis based on their specific needs and expertise levels."

Testing the Framework: From Sepsis to Kidney Disease

To demonstrate the framework's capabilities, the researchers conducted proof-of-concept tests using data from the PhysioNet Challenge 2019, which focused on early sepsis prediction, and the Stockholm Creatinine Measurements (SCREAM) database for kidney disease analysis.

The tests examined four different scenarios:

1. Infection Progression Analysis: Tracking how patients move through different states of infection and inflammation, from normal temperature to sepsis.

2. Organ Damage Progression: Analyzing how organ dysfunction develops across cardiovascular, renal, and hepatic systems.

3. Kidney Disease Progression (Moderate): Following the trajectory of patients with chronic kidney disease through different stages of declining function.

4. Kidney Disease Progression (Severe): Examining more severe cases requiring kidney replacement therapy.

The framework successfully processed all test datasets and generated detailed reports for each scenario. The process maps revealed critical patterns, such as temperature fluctuations as early indicators of sepsis and cardiac damage as a gateway to multi-organ dysfunction.

"What's particularly valuable is how the system identifies not just what happened, but when interventions might be most effective," explains Farhad Abtahi, another researcher on the project. "For example, we found a 6-7 hour window for intervention based on temperature patterns, which could be crucial for clinical decision-making."

AI Models: Not All Created Equal

One of the most interesting findings was the variation in performance among different AI language models. The researchers tested five state-of-the-art models: Claude Sonnet-4 (Anthropic), GPT-4.1 (OpenAI), Gemini 2.5 Pro (Google), DeepSeek R1, and Grok-4 (X-AI).

Claude Sonnet-4 and Gemini 2.5 Pro emerged as the clear leaders, with the highest overall scores for accuracy and consistency. Claude achieved an impressive average score of 3.79 out of 4.0 across all test cases, while Gemini followed closely at 3.65.

"We were particularly impressed with Gemini 2.5 Pro's performance," notes Illueca Fernandez. "It was the only model that consistently avoided hallucination—generating information not present in the data—which is critical for clinical applications."

The researchers also conducted a cost-effectiveness analysis, revealing substantial differences in efficiency. DeepSeek R1 achieved the highest performance-to-cost ratio, costing just $0.02 per report while maintaining reasonable accuracy. In contrast, GPT-4.1 was the most expensive at $1.13 per report.

"Cost considerations are important for real-world implementation," explains Chen. "Our analysis shows that choosing the right model can reduce costs by up to 98% without significantly compromising quality."

Clinical Implications: From Data to Action

The framework's ability to translate complex process mining results into clinically relevant insights could have significant implications for healthcare delivery. The researchers identified several potential applications:

• Early Intervention Protocols: Temperature pattern analysis could help identify sepsis risk 6-7 hours before clinical manifestation.

• Multi-organ Monitoring: Cardiac biomarkers might serve as early warning signs for cascading organ dysfunction.

• Medication Risk Assessment: Comparing proton pump inhibitors (PPIs) versus H2 blockers (H2Bs) revealed faster kidney disease progression with PPIs.

• Risk Stratification: The system identified a 2.7-9x higher risk of kidney function decline in patients taking PPIs compared to H2Bs.

"These findings demonstrate how process mining can move beyond descriptive analysis to generate actionable insights for clinical practice," says Seoane. "The framework doesn't just tell you what happened—it helps you understand why it happened and what you might do about it."

Limitations and Future Directions

While the results are promising, the researchers acknowledge several limitations. The current study represents a technical proof-of-concept rather than a clinically validated tool. The evaluation used synthetic and retrospective data, and the outputs have not yet been validated by healthcare professionals.

"This is just the beginning," emphasizes Abtahi. "Our immediate priorities include usability testing with 20-30 healthcare professionals and comparing AI-generated reports against clinician interpretations."

Future work will focus on prospective deployment in clinical settings, expanding validation beyond sepsis and kidney disease, and incorporating real-time data streams for continuous process monitoring.

"Until these validation studies are complete, HealthProcessAI should be considered a research tool for exploring process mining applications rather than a clinical decision support system," cautions Illueca Fernandez.

A New Approach to Healthcare Analytics

HealthProcessAI represents a significant advancement in making complex healthcare analytics more accessible and useful. By combining the analytical power of process mining with the interpretive capabilities of AI language models, the framework addresses key barriers to adoption.

"What sets our approach apart is the deliberate integration of educational scaffolding with advanced AI capabilities," explains Chen. "Unlike existing solutions that typically address either technical process mining or clinical AI in isolation, HealthProcessAI bridges both domains with integrated educational support."

The framework's modular architecture allows for customization based on specific clinical needs and expertise levels. Its multi-model orchestration approach leverages the strengths of different AI systems, enhancing interpretive accuracy and cost efficiency.

"As healthcare continues to digitalize, tools that can make sense of complex data while remaining accessible to non-technical users will become increasingly valuable," says Seoane. "HealthProcessAI offers a validated, accessible, and scalable approach to advancing clinical process intelligence."

The researchers have made all framework components, documentation, and sample datasets available through the HealthProcessAI GitHub repository under an MIT license, encouraging further development and adaptation by the healthcare community.

As healthcare systems worldwide grapple with increasing complexity and data overload, frameworks like HealthProcessAI could play a crucial role in transforming raw information into meaningful insights that improve patient care and outcomes.