Talk "Process Transparency Through Data Mining: A Hospital Case Study"

Topic was presented at Munich Datageeks - December Edition 2025

Abstract

This talk presents a case study on process transparency and process mining in a hospital setting, specifically focusing on the perioperative process at Großhadern, one of Germany's largest hospitals. The speaker, Harald from QAware, demonstrates how process mining can provide valuable insights into hospital operations without requiring complex data analysis. The study reveals that healthcare expenditures, particularly surgical procedures, represent significant costs that could be optimized through better process understanding. Using event log data from Sqior Medical's patient logistics software, the team discovered that simple statistical methods like arithmetic mean often outperform sophisticated AI models in predicting operation durations. Key findings include the critical importance of domain expertise in data interpretation, the prevalence of manual planning inefficiencies, and the value of real-time process visualization for empowering hospital staff. The talk emphasizes that augmentation rather than automation of human expertise delivers better results, and that simpler solutions are often more effective than complex AI-driven approaches.

About the Speaker

Harald is a senior consultant at QAware, a boutique software engineering company based in Munich with approximately 250 employees. With extensive experience across multiple affiliations and nearly a decade at QAware, he bridges the gap between academic research and practical software engineering. His work focuses on high-end, mission-critical software projects where failure is not an option, particularly in applying transformative technologies like cloud computing and AI to business-critical systems. QAware maintains a stake in Sqior Medical, providing software engineering services to the healthcare startup, which led to Harald's involvement in this process mining project.

Summary

Healthcare Context and Motivation

Healthcare expenditures have increased dramatically across Western countries over the past 40 years. In Germany, healthcare spending has risen from 8.1% of GDP in 1980 to over 12% by 2020, representing over 500 billion euros annually as of 2024. Hospital treatments account for the largest portion of these expenditures at 33%, and while the percentage growth has slowed recently, absolute spending continues to rise year after year. This creates a compelling economic case for optimizing hospital processes.

The Großhadern Hospital Case

Großhadern represents one of Germany's largest hospitals with 5,500 staff members handling 300,000 cases annually. Surgical procedures are particularly significant, accounting for 40% of hospital expenses while generating 60% of revenue. Even modest improvements in the perioperative process could yield substantial cost savings. The perioperative process encompasses the entire workflow around surgery, from patient admission through pre-operative preparation, anesthesia, the actual surgical procedure, recovery, and discharge.

Technology Partnership

Sqior Medical, a Munich-based startup in which QAware holds a stake, provides IT infrastructure and software for patient logistics in hospitals. They serve many of Germany's largest hospitals, including university clinics. QAware's participation involves providing software engineering services rather than financial investment. Sqior Medical approached QAware to analyze their data and implement process mining capabilities to enhance their already successful solution.

Process Mining Fundamentals

The talk distinguishes between several related concepts:

Classic process mining involves analyzing historical event logs from IT systems to understand what happened in the past. This one-time analysis of data from several months prior can deliver valuable insights and represents the core business of companies like Celonis, one of Germany's unicorn companies.

Process observability goes beyond historical analysis by processing data in real-time on a continuous basis. This enables interactive visualization that empowers process stakeholders like nurses and medical staff to see what is happening as it unfolds. The visualization itself often enables people to identify and resolve issues before automated anomaly detection systems activate, representing an optimal form of process improvement.

Data science contributes the analytical component, testing specific KPIs and hypotheses. Business process management focuses on people and tasks, creating detailed process models through extensive organizational analysis. QAware's business approach uses these insights as a vehicle to improve the underlying IT systems rather than just the processes themselves.

The Reality Gap

In theory, processes consist of sequential steps supported by people and systems like Salesforce or SAP. Business process management model the people and processes, while companies like Celonis take the system approach extracting data from systems. However, reality proves more complicated. Many process steps exist on paper, in Excel spreadsheets, or remain undocumented. Legacy systems running decades-old software often lack logs or easy data access. Additional stakeholders may be involved who are unknown or whose data cannot be accessed. Achieving complete process transparency requires examining both people and all systems, not just the major enterprise platforms.

Initial Observations

Harald conducted on-site interviews at Großhadern with operation planners and doctors, revealing several key findings:

The perioperative process is not uniformly understood. Different stakeholders name steps differently and disagree on responsibility boundaries. Creating a common process model became essential just to enable coherent communication.

The process does manifest in the data through a sequence of steps covering induction, preparation, procedure, and recovery. However, the data probe names bear no relation to what people call these steps in practice. Fortunately, specific data points could be mapped to process steps, though this alignment is not always straightforward.

Process Discovery Challenges

Using standard process discovery tools like PM4Py on the raw event logs produces a single giant net that folds all processes together. This visualization proves too complex to extract meaningful insights. Refining and condensing this output to extract individual flows requires significant domain understanding. While analysts working repeatedly in standard domains like SAP procurement processes build this knowledge over time, entering a new domain necessitates acquiring domain expertise first.

Visualization and KPIs

Once properly refined, the visualization displays process flows with varying lane widths representing patient volumes through different pathways. Simply showing hospital staff this real-time visualization of their current state enables them to derive helpful insights, making the visualization immediately valuable even before deeper analysis.

The system enables tracking individual patients and continuously updating key performance indicators. One crucial KPI is plan stability. Hospitals must maximize operating theater utilization since significant capital is tied up in these facilities. However, they cannot overbook because postponing a patient's surgery to the next day means the hospital cannot obtain reimbursement from health insurance for that day's bed occupancy. Conversely, leaving operating rooms underutilized wastes expensive resources. Additionally, pushing surgeries into overtime strains doctors, staff, and is suboptimal for patients. Achieving high plan stability, where operations consistently occur within their scheduled time windows, represents a critical optimization goal.

Influencing Factors

The team identified numerous factors potentially affecting operation duration, including:

• Duration of specific phases: induction, preparation, procedure, and recovery
• Patient characteristics: age affects recovery time differently for young children, adults, and elderly patients
• Technical factors: patient positioning on the operating table can be simple or complex depending on the surgical site
• Procedure type and complexity

Notably, the team deliberately excluded contextual factors like specific hospitals, clinics, and individual surgeons from analysis. While this data was available and doctors actually wanted it analyzed, the team refused because analyzing individual surgeon performance would involve sensitive personal data. This decision ultimately strengthened rather than weakened the analysis.

Data Quality Challenges

The dataset covered one full year but yielded only 11,000 usable rows describing procedures with anesthesia. This limited data quality stems from poor compliance with manual data entry requirements. Only automatically gathered data proves reliable for analysis. Furthermore, the dataset contained several hundred different procedure types, creating very small sample sizes for many categories.

The procedure type field consisted of free text rather than structured categories. This resulted in over 2,000 distinct text descriptions across 11,000 procedures. Variations included typos, different abbreviations, and alternative phrasings like leg amputation versus amputation leg. Automated clustering produced approximately 135 procedure categories, but clinical experts identified only 15 truly distinct procedure types, just over 10% of the automated result. Only about half of these proved practically meaningful for analysis.

The Clustering Problem

Specific examples illustrate the challenge. Some staff write MS while others write Magensonde (the German term). The Levenshtein distance between these terms is large, making automated similarity matching ineffective. Similarly, variations in German anesthesia terms like Art (wach), Wach-Art, Wach.Art, and Wachaterie demonstrate how automated similarity measures fail without domain knowledge. This clustering task requires expert knowledge and cannot be automated with current technology. Expert knowledge proves crucial for meaningful data analysis, demonstrating that process mining involves people as much as technology.

Planning Analysis

Comparing actual operation durations against manually created plans revealed significant deviations. A scatter plot showing actual duration on the x-axis versus planned duration on the y-axis demonstrated that very few cases fell within a plus-or-minus 20% acceptable range along the diagonal. More tellingly, horizontal banding patterns appeared in the distribution, revealing that planners round to 15-minute increments. A frequency distribution confirmed this, with planning data showing peaks every 15 minutes while actual durations followed a smooth normal distribution.

Human planning behavior shows systematic bias toward safety margins. Actual durations prove substantially shorter than planned durations, reflecting the understandable tendency to plan conservatively. However, this conservative approach undermines plan stability by systematically overestimating time requirements.

Predictive Modeling Attempts

A master's student tested every major predictive method including random forests and various AI approaches to forecast operation durations based on the identified factor model incorporating patient characteristics, procedure type, and other variables. The results proved surprising: the simple arithmetic mean performed as well as or better than sophisticated machine learning models in most cases. The arithmetic mean offers several advantages over complex models. It can be computed in real-time, updating every millisecond as new data arrives, whereas complex models require lengthy retraining periods. Its simplicity makes it transparent and maintainable.

The research team repeatedly verified these results because they seemed too simple to believe. The doctors, however, found the team's surprise puzzling, apparently finding the average method's effectiveness unsurprising.

Key Insights and Takeaways

Process transparency encompasses far more than conventional process mining tool application. Real-world analysis requires understanding at the front end, tool-based data extraction and processing in the middle, and deriving actionable insights at the conclusion.

Combining qualitative and quantitative analysis produces superior results. Automated machinery alone struggles with this synthesis, suggesting job security for human analysts. The high variation in free-text descriptions and strong context dependency resist automation and break AI approaches. The recommendation is to pursue augmentation rather than automation, improving human practice with tools rather than attempting to replace human expertise entirely.

Process mining serves as an excellent stepping stone for meaningful data analysis that provides client value. Low-hanging fruit abounds for those willing to pursue it. AI does not answer every question, despite its power as a tool. The current AI hype appears to be descending the downward slope of the hype cycle as practitioners recognize its limitations alongside its capabilities. The fundamental principle that simpler solutions are better proves increasingly relevant. While complex solutions appeal to computer scientists and programmers who enjoy over-engineering, creating custom databases, frameworks, and programming languages, simplicity delivers better practical results.