SIAP: an intelligent algorithm for multiple prescription pattern recognition based on weighted similarity distances

Yifei Wang, Julia Xu, Jie Zhang, Hong Xu, Yuzhong Sun, Yuan Miao, Tiancai WenBMC Medical Informatics Decis. Mak 2023 · DOI

TL;DR

This paper introduces SIAP, a novel algorithm that intelligently identifies and classifies complex drug prescriptions by weighting individual drug importance. This approach significantly improves upon existing methods for recognizing prescription patterns, offering a more nuanced understanding of drug combinations and their therapeutic roles, particularly in Traditional Chinese Medicine (TCM).

What Problem Does This Paper Address?

Clinical practice, especially for chronic diseases, often involves complex prescriptions with multiple drugs. Identifying the specific role and importance of each drug within these combinations is challenging due to variations in patient factors, disease complexity, and the sheer number of potential drug interactions. Existing methods for prescription pattern recognition, like the Intersection Set Rate (ISR) algorithm, are limited as they treat all drugs equally and cannot effectively handle the multidimensionality of complex prescriptions. This gap hinders the extraction of valuable clinical knowledge and the development of optimized treatment strategies.

Methods

The core of this research is the Subgroup Identification Algorithm for Complex Prescriptions (SIAP). SIAP introduces a weighted approach to prescription analysis:

  • Drug Weighting: Drugs within a prescription are categorized into four levels (C1-C4) based on their importance, with a fifth category (C5) encompassing all drugs. These weights are derived from real-world clinical data and expert consensus, reflecting the hierarchical importance of drugs in treatment.
  • Weighted Similarity Distance: SIAP calculates a similarity coefficient (sim) by combining distance coefficients (d) between the prescription to be identified and candidate standard prescriptions, weighted by the importance of each drug category (w). This moves beyond simple set intersection to a more nuanced comparison.
  • Logistic Regression Model: A logistic regression model is employed to estimate the relative importance weights (w) of the different drug categories.
  • Candidate Prescription Generation: A simple pooling algorithm identifies potential standard prescriptions that share at least one drug with the input prescription.
  • Threshold Optimization: Instead of a fixed threshold, SIAP sequentially tests similarity thresholds and selects the one that maximizes the F1 score, a balanced measure of precision and recall.
  • Comparison Algorithms: SIAP was compared against the baseline ISR algorithm and two variants of SIAP: SIAP-All (using C1-C4 weights) and SIAP+All (using C1-C5 weights).

Datasets

  • Standard Prescription Dataset: 376 traditional Chinese medicine (TCM) formulas were collected from established texts and websites, with herbs classified into C1-C4 levels.
  • Clinical Prescription Dataset: 1438 real-world TCM prescriptions were extracted from the TCM Data Center of the China Academy of Chinese Medical Sciences. These were manually identified by experienced TCM doctors and used for training and testing.
  • Training/Test Split: The clinical dataset was split into a training set (70%, n=1000) and a test set (30%, n=438).

Key Insights and Ideas

  • Drug Importance Matters: The most significant insight is that not all drugs in a prescription are equal. Assigning weights based on clinical importance dramatically improves the accuracy of prescription pattern recognition.
  • Beyond Simple Matching: SIAP's weighted similarity approach provides a more sophisticated way to match complex prescriptions to known standards, moving beyond the binary "yes/no" of simpler methods.
  • TCM as a Proving Ground: The complexity of TCM prescriptions, with their multi-herb combinations and layered therapeutic principles, makes it an ideal domain to test and demonstrate the effectiveness of such a weighted algorithm.
  • Bridging Clinical Experience and Data: The algorithm effectively translates expert knowledge (drug importance) and clinical data into a computational model, facilitating knowledge extraction and potentially aiding less experienced practitioners.

Key Results

  • Superior Performance: SIAP-All and SIAP+All significantly outperformed the baseline ISR algorithm.
  • F1 Score Improvements: SIAP+All achieved an F1 score of 0.7799, an 11.04% improvement over ISR. SIAP-All achieved an F1 score of 0.7568, an 8.73% improvement.
  • Precision and Recall: SIAP variants showed notable improvements in precision, with SIAP+All reaching 0.9430, while maintaining competitive recall.
  • Software Development: A functional TCM prescription identification software was developed based on the SIAP+All algorithm, demonstrating its practical applicability.

Limitations and Future Work

  • Limited Parameters: The current study did not include information on underlying diseases, symptoms, signs, or patient living areas, which could further refine the model.
  • Future Expansion: The authors suggest expanding the parameters included in the model and increasing the validation data to optimize its performance.
  • Generalizability: While tested on TCM, further work could explore the applicability of SIAP to Western medicine prescriptions or other complex medical data.
  • Dose and Frequency: The current study did not consider drug dosage or frequency, which are crucial aspects of actual clinical practice.

Summary updated

BibTeX
@article{wang2023siap,
  author = {Yifei Wang and Julia Xu and Jie Zhang and Hong Xu and Yuzhong Sun and Yuan Miao and Tiancai Wen},
  title = {{SIAP: an intelligent algorithm for multiple prescription pattern recognition based on weighted similarity distances}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2023},
  doi = {10.1186/s12911-023-02141-3},
}
Figure