gms | German Medical Science

Background: A contemporary pharmacotherapy is based on randomized controlled trials and guidelines. General practitioners and pharmacists often feel overwhelmed by rapid changes in therapeutic standards. As a consequence, guideline adherence is very limited even in widespread diseases as hypertension, dyslipidemia, coronary heart disease or chronic heart failure [1], [2], [3]. Clinical decision-support systems (CDSS) are defined as a software that is designed to be a direct aid to clinical decision-making and are widely accepted in current medical and pharmaceutical practice [4], [5]. Artificial intelligence (AI) is used in medicine predominately in diagnosis or whenever a large amount of data needs to be analyzed [6]. In contrast to its expected application in diagnosis, AI is rarely employed in therapy. This might be due to the complexity of patients and individual complaints. Theoretic approaches to CDSS in pharmacotherapy are to provide a fixed, algorithm-based software or, in contrast, to learn from real patient scenarios. An algorithm-based scenario may face restrictions in multimorbid patients with contradicting therapies. AI on the other hand can learn from clinical practice but would reflect the current level of practice if not guided by specialists. The aim of this study was to study and compare different AI methods for a CDSS in pharmacotherapy.

Materials and methods: The data used for the study was taken from a previous clinical study on Medication Management, in which clinical pharmacists and general practitioners (GPs) collaboratively optimized the pharmacotherapy of community patients for at least 12 months. The published study protocol and the study results provide a detailed description of the WestGEM study [7], [8]. For AI modeling, data of 76 patients was drawn from the patient cohort of 142 patients and transferred into the software R Studio (version 1.2.1335, R foundation, Austria) for further modeling. As specific models, Decision Tree, Random Forest and Neural Network were tested.

Results: Data of 76 patients with 75 drug classes in use was generated. Median age was 77 years, the average number of diagnoses was 5.8 and the average number of drugs was 8.5. Leading diagnoses were hypertension (86.6% of patients), coronary artery disease and hyperlipidemia (43.4%). Median age was 77 years, the average number of diagnoses was 5.8 and the average number of drugs was 8.5. Leading diagnoses were hypertension (86.6% of patients), coronary artery disease and hyperlipidemia (43.4%). Drugs which appeared in less than 10 percent of the patients were excluded. This let to 23 drugs which were used for further modeling.

Decision Tree: The Decision Tree did not produce any results for drugs, which were seen in ≤17% of the patients. A similar problem happened for very frequent drugs (i.e. ACE-Inhibitors).

Random Forest: As the Random Forest uses the Decision Tree as a foundation, no results were seen for drugs, which were used by ≤17% of the patients as well.

Neural Network: The Neural Network produced decisions for all drugs but was prone to overfitting.

Conclusion: The three analyzed AI approaches showed different characteristics. Results of the Decision Tree can be followed and displayed clearly so that they can be discussed. This was a great advantage of this model. The Random Forest showed a similar susceptibility to errors. The absence of errors in the Neural Network was interpreted as an indicator for overfitting. As a conclusion of this basic assessment of AI approaches in pharmacotherapy, it was realized that none of the tested methods produced helpful results at this stage of development. More research is needed.