From Manual Search to Artificial Intelligence: A Methodological Framework for Developing Patient Journey Models from Electronic Health Records Using Natural Language Processing

Background: Improving care pathways for chronic conditions requires a holistic, data-driven understanding of the patient journey/ care continuum. This study introduces a novel framework to reconstruct and simulate patient journeys using medical data and probabilistic modeling. Chronic spontaneous urticaria (CSU) was selected as a use case, with first documented symptom occurrence estimated through natural language processing (NLP) and treatment cycles identified from Anatomical Therapeutic Chemical (ATC) code sequences, an approach not previously applied in this context. Objectives: To develop and evaluate a potentially scalable framework for reconstructing patient journeys and simulating care pathways in chronic diseases, with the aim of identifying and quantifying care gaps and finding opportunities to reduce delays, improve treatment initiation, and support value-based care. Methods: A Markov model was developed to simulate transitions across health states using de-identified electronic health record (EHR) data from 74,664 patients with CSU. Patient journeys were reconstructed by combining NLP-based estimation of first documented symptom occurrence with medication treatment cycles derived from ATC code sequences. The model was then used to characterize the current care pathway. Results: The analysis identified substantial delays in diagnosis and considerable variability in treatment initiation. Incorporation of NLP-derived symptom information from unstructured clinical notes increased the number of patients with a documented symptom record from 29,457 to 56,551. The interval between first documented symptom and diagnosis was 157 days (median 0) using ICD-9/10 codes only and 570 days (median 624) using ICD-9/10 codes plus NLP. Because the NLP output was not independently validated and the look-back period was limited, these values are estimates and should be read with caution. Notably, 29.7% of patients received no prescribed therapy. Conclusions: This study demonstrates the feasibility of reconstructing and simulating patient journeys using routinely collected healthcare data. Although intentionally simple, the conceptual framework may be applicable across chronic conditions, however, its implementation would require disease-specific adaptation and validation. The Markov model provides a framework for simulating targeted interventions designed to streamline care pathways, reduce delays, and optimize resource utilization in order to support data-driven decision-making, selection of targeted interventions, care pathway optimization, and the advancement of value-based healthcare.