Clinical Knowledge Model for Lung Disease Exposes Implicit Practice and Disagreements

A digital clinical knowledge model designed to guide self-management in chronic obstructive pulmonary disease revealed key implicit clinical practices and substantial interprofessional disagreements when evaluated against fictional patient scenarios, researchers report in BMC Digital Health. Clinical knowledge models encode explicit medical rules and decision logic to automate clinical reasoning and patient decision support, yet formalized clinical practice guidelines rarely capture the full breadth of real-world care routines. To determine whether piloting a rule-based digital model with fictional scenarios could elicit unrecorded medical expertise and evaluate clinical consensus, a research team developed a prototype decision system and tested its recommendations against the judgments of practicing respiratory clinicians.

Healthcare delivery frequently demands that clinicians execute networks of interrelated decisions simultaneously, such as determining discharge readiness, triage classifications, and pharmacological adjustments. While formal guidelines outline standard management pathways, practicing clinicians routinely rely on implicit knowledge gained through personal clinical experience that is never formally documented. The investigators constructed a clinical knowledge model structured around digital concepts, associated attributes, and decision tables containing rules that continuously process incoming patient variables. This transparent architecture differs fundamentally from data-driven machine learning models, as every generated recommendation can be traced directly back to specific executing rules and physiological inputs, avoiding the black-box opacity that often causes clinicians to distrust automated advice.

The system evaluated in the study was developed to assist patients diagnosed with chronic obstructive pulmonary disease in recognizing and self-managing impending acute flare-ups, thereby reducing routine monitoring burdens on specialized hospital departments. The application evaluated daily inputs consisting of three vital physiological parameters—oxygen saturation, heart rate, and respiratory rate—alongside responses to the COPD Exacerbation Recognition Tool. When symptom scores indicated potential clinical deterioration, the software triggered two supplementary questionnaires: the Exacerbations of Chronic Pulmonary Disease Tool to evaluate the duration and severity of coughing and breathlessness, and a custom questionnaire addressing guideline-mandated symptoms including chest pain, fever, hemoptysis, palpitations, orthopnea, and ankle edema.

To evaluate the model across a wide range of clinical variables without exposing actual patients to unvalidated software, the researchers designed six fully synthetic patient cases using a knowledge-driven methodology developed in three collaborative sessions with an independent pulmonologist. Each synthetic profile detailed baseline and current vital signs, patient age, prior medical history, and completed symptom questionnaires reflecting stable chronic obstructive pulmonary disease under home monitoring. Eight healthcare professionals recruited from the Department of Respiratory Medicine at Maastricht University Medical Center+ participated in the evaluation, comprising four pulmonology specialists, two pulmonologists in training, and two non-physicians with direct patient contact, specifically a nurse practitioner and a physician assistant, possessing a median professional experience of eleven years.

During individual in-depth interviews, the clinicians reviewed the first three synthetic cases in an open format to generate their own unprompted clinical management plans, allowing researchers to measure deviations from the model’s automated outputs. Across these three initial scenarios, inductive coding of the transcripts revealed six high-level clinical themes and twenty-four distinct response categories: retaking physiological measurements, asking supplementary clinical questions, initiating care professional contact, modifying pharmacological therapy, continuing home monitoring, and offering non-pharmacological interventions. Clinicians systematically sought additional context that the automated system did not capture, posing follow-up inquiries regarding outdoor exposures, daily household functionality, animal contact, and deviations from baseline respiratory symptoms in 20 out of 24 case evaluations.

The comparative analysis demonstrated that clinicians approached patient instability with greater caution than the decision model, frequently proposing more aggressive interventions. In the first test scenario, seven out of eight clinicians advised repeating vital measurements to rule out acute physical exertion or erroneous reporting, suggesting retest intervals ranging from five minutes to one hour, whereas the software prescribed an immediate retest after two minutes. In more severe exacerbation scenarios, the automated model advised increasing existing inhaler dosages and adopting non-pharmacological measures such as airway clearance techniques and paced breathing. In contrast, several clinicians bypassed dosage escalations in favor of prescribing systemic antibiotics or initiating new drug regimens, while completely omitting the lifestyle and breathing recommendations provided by the digital tool.

Quantitative scoring of discrepancies revealed that the degree of divergence between human clinicians and the automated model peaked in case two, which registered twenty-nine total thematic deviations and a mean deviation score of 3.6. Furthermore, comparing clinician responses to one another exposed prominent interprofessional differences across both different clinical roles and within identical specialties. Heatmap analysis demonstrated average inter-clinician difference scores of 3.4 for case one, 7.3 for case two, and 5.6 for case three. While fully certified pulmonologists demonstrated higher internal consistency than comparisons with other roles, non-physician providers exhibited the highest internal divergence, highlighting that clinical decision-making strategies vary widely even within a single specialized hospital department.

When presented with the remaining three synthetic scenarios alongside the model’s generated recommendations, clinician acceptance varied dramatically by case context and practitioner seniority. All eight clinicians fully adopted the automated advice for case four, viewing disagreements over recommended monitoring intervals as minor. However, acceptance dropped to exactly fifty percent for case five due to conflicting views on inhaler dosage modifications, and fell to twelve percent in case six, where seven of the eight clinicians rejected the automated guidance. Clinicians criticized the case six output for overwhelming patients with excessive detail, noting that advising patients to simultaneously schedule laboratory diagnostics and contact their specialist undermined the physician’s autonomy to determine necessary diagnostic workups.

Clinician adoption behavior correlated directly with age and professional experience, revealing an unexpected trend that contradicts prevailing literature on medical automation bias and technology acceptance. While younger clinicians and trainees frequently rejected the automated advice, older and more senior specialists accepted the model’s recommendations at substantially higher rates, with acceptance rising steadily alongside participant age. Conventional medical informatics literature suggests that less experienced practitioners tend to rely more heavily on automated decision aids, while experienced physicians often exhibit skepticism toward algorithmic recommendations that conflict with intuitive clinical judgment. The authors noted that further research is required to untangle whether chronological age or cumulative practice experience drives this increased receptivity.

The study authors acknowledged significant methodological limitations, including a small sample size drawn from a single academic medical center and the intentional exclusion of real patients from this preliminary verification phase. Because the digital model was tailored specifically to institutional practices at Maastricht University Medical Center+, the observed consensus hurdles may differ across broader healthcare settings. Future evaluations must involve larger multi-center clinician cohorts and evaluate live patient interactions to verify usability and real-world safety. Ultimately, the investigation demonstrates that knowledge-driven digital models can systematically extract hidden clinical practices and expose critical interprofessional divides, establishing essential baseline requirements before automated decision support tools can be safely deployed into routine pulmonary care.

The investigators emphasized that the clinicians’ tendency to request extensive supplementary details and prescribe stronger therapies reflects broader habits in respiratory practice, where overtreatment of chronic obstructive pulmonary disease is estimated to occur in nearly one in two cases. By codifying strict decision criteria, a clinical knowledge model could potentially reduce overtreatment by giving practitioners the confidence to work with less information and seek additional details only when strictly necessary for decision-making. Furthermore, such digital systems could prompt clinicians to utilize critical diagnostic parameters that are frequently neglected in routine clinical workflows, such as spirometry during initial patient evaluations to prevent erroneous diagnoses.

The study also situated its methodology within the broader landscape of digital pulmonary tools, comparing its human-centered evaluation against earlier decision-support systems. Whereas previous systems such as the ACCESS prediction model relied on statistical receiver operating characteristic analyses to demonstrate predictive performance, other platforms like COPDPredict established validity by directly comparing algorithm-defined exacerbations with clinician-defined events across both patient and physician cohorts. The authors mirrored this comparative approach to verify whether their model’s automated outputs conformed directly to actual clinical practice, aiming to prevent declines in care quality as automated decision tools become more prevalent across health systems.

To contextualize the substantial interprofessional disagreements observed during the case evaluations, the authors drew upon medical decision-making frameworks identified by Svensson and Jacobsson in emergency departments. That framework categorizes clinical decisions into experiential, ostensive or rule-based, and action-based modes, which interact to form an eight-fold typology of clinical practice. The researchers noted that these divergent decision-making styles indicate that interprofessional variation is not unique to respiratory specialists, and the interview materials generated during the study offer valuable insight for clinical teams seeking to deliberate and reconcile differing management styles even outside digital health implementation.

Because transparent clinical knowledge models depend entirely on predefined domain rules and cannot independently generate new rules from accumulated datasets, the authors proposed addressing this structural limitation through a hybrid two-agent architecture. Under this proposed closed-loop design, the deterministic rule engine would operate alongside a complementary data-driven machine learning component that observes automated model recommendations, recorded clinician interventions, and resulting patient outcomes. This secondary component could inductively synthesize novel decision rules over time while preserving the traceable logic and explainability required to maintain trust among healthcare professionals and patients.

Subject of Research: Evaluation of a digital clinical knowledge model for chronic obstructive pulmonary disease self-management using synthetic cases and expert interviews

Article Title: Piloting a clinical knowledge model: digital modeling to support the self-management of patients with chronic obstructive pulmonary disease

Article References: Edelman, E., Bellemakers, E., Tijssen, F., Munniksma, P. R., Bast, W., ten Bohmer, H., Simons, S. O., Spreeuwenberg, M., & van Merode, F. (2026). Piloting a clinical knowledge model: digital modeling to support the self-management of patients with chronic obstructive pulmonary disease. BMC Digital Health, 4(1), Article 44. https://doi.org/10.1186/s44247-026-00278-5

Image Credits: AI Generated

DOI: 10.1186/s44247-026-00278-5

Keywords: chronic obstructive pulmonary disease, clinical knowledge model, clinical decision support, digital health, knowledge elicitation, interprofessional variation, medical informatics, pulmonology, Piloting, clinical, knowledge, model