An intelligent algorithm based on weighted similarity distances for multiple contributing prescription recognition: SIAP

Abstract Background. Patients with chronic diseases, especially the elderly, often suffer from more than one disease in clinical practice. Different diseases have different clinical manifestations for different patients. Doctors normally will identify the primary and the secondary diseases, and then organise the corresponding medication, their dosage and duration that suit the patient best. The complex considerations present a challenge in standardization and optimization of prescriptions. What have been recorded in the medical record systems are just all the medicine prescribed, without information on the medicines for individual diseases. Therefore, being able to analyze the medical records and identify the composition of the medicines for different diseases is essential for further analysis and standardization. Methods. In this study, we developed a Subgroup Identification Algorithm for complex Prescriptions, SIAP. The algorithm can identify the‘sub’prescriptions for each diseases from a complex prescription for all the diseases. The approach firstly established a standard prescription database based on prescriptions in guidelines and expert consensus. Each disease has one or several standard prescriptions. Each prescription has a number medicines, each with a weight according to its importance. Given a complex prescription, SIAP matches its subsets to the standard prescription database with more weights assigned to medicines of higher priority, and iterate the algorithm to ind the optimal match. We collected 376 standard prescriptions (formulas) for the corresponding diseases. We further collected 1438 complex prescriptions from the clinic record system. SIAP has been verified over the data collected, in two variants of the algorithm, SIAP-All and SIAP+All. The accuracy, recall and F1 values were used to judge the merits of SIAP-All and SIAP+All algorithms, and compared to the baseline algorithm ISR. Results. Both of the SIAP-All and SIAP+All algorithms outperformed the benchmark algorithm ISR in terms of accuracy, recall, and F1 value. Particularly, the F1 values are 0.7568 for SIAP-All and 0.7799 for SIAP+All, improved 8.73% and 11.04% correspondingly over the benchmark algorithm ISR. Conclusion. The results show that the SIAP significantly outperforms the benchmark algorithm ISR. This research can help to identify and analyze medicine combinations in complex prescriptions, laid a foundation for its optimization or standardization.