RP14 - Clinical pathway discovery based on electronic health record data

Clinical guidelines and their corresponding standard operating procedures (SOPs) are extensively used sources of knowledge in the treatment of melanoma patients offering evidence-based information valuable for informed medical decision-making. Electronic health records (EHRs) provide vast amounts of real-time patient care data. This presents an opportunity to develop data-driven models, methods, and tools to connect SOPs and clinical pathways with patient context.

However, today´s hospital information systems often lack electronic patient documentation that is primarily oriented towards clinical needs. Consequently, physicians and surgeons may need to search through a significant amount of data to locate currently relevant patient information for the next treatment decision, rather than having direct access to the relevant information during the current treatment step.

To address these challenges, our project is based on results from prior projects focussing on incorporating patient data into clinical pathways [1] to simulate patients’ way through a given pathway and enhancing the data quality and usability of EHR data for clinical applications (IWG data quality). Furthermore, machine learning methods have been successfully used to learn the most common clinical paths for groups of patients based on Electronic Health Records [2], [3] and will be applied here.

To dynamically learn the best next step and provide decision support for an individual patient based on their personal medical data an approach to extract event logs from clinical health record data using process mining techniques will be developed. These event logs could be visualized as BPMN-models and serve as a patient-specific clinical pathway representing the patients’ disease progressions and treatment histories. With the ability to locate a patient’s state in a clinical pathway, the SOP- which describes the ideal pathway and is based on evidence-based guidelines and recommendations - could be compared with the learned one to determine the next steps in the treatment process on a patient-by-patient basis.

To discover pathways from data we will evaluate stochastic methods such as Hidden Markov models used by Zhang et al [2] and Meier et al [3] as well as deep-learning-based methods like Variational and Predictive AutoEncoders (VPAEs).

The results can potentially serve as a foundation for future projects on simulating treatment outcomes based on a given patient’s unique circumstances and medical history.

[1] Catharina Lena Beckmann et al. “Guideline-Based Context-Sensitive Decision Modeling for Melanoma Patients”. In: Studies in Health Technology and Informatics 296 (Aug. 17, 2022), pp. 50–57. doi: 10.3233/SHTI220803.

[2] Yiye Zhang, Rema Padman, and Nirav Patel. “Paving the COWpath: Learning and visualizing clinical pathways from electronic health record data”. In: Journal of Biomedical Informatics 58 (Sept. 28, 2015), pp. 186–197. doi: 10.1016/j.jbi.2015.09.009.

[3] Jens Meier et al. “Predicting treatment process steps from events”. In: Journal of Biomedical Informatics 53 (Dec. 12, 2014), pp. 308–319. doi: 10.1016/j.jbi.2014.12.003.