考量管制藥特性之機器學習需求預測研究-以鴉片類藥物為例

Abstract

鴉片類藥物(Opioids)經常被使用在癌症以及慢性病的治療中，能夠有效的緩解疼痛，而隨著癌症以及慢性病人口的增加以及發病年齡層下降的影響，鴉片類藥物需求量也隨年增加。然而，鴉片類藥物卻存在依賴、成癮和過量的風險，甚至有心人士會不當轉移(diversion)此類藥物，進而合成類似海洛因(Heroin)的毒品，危害社會。基於上述原因，若能有準確的需求預測值，便能作為成為辨別異常開立處方箋的一個基準線，協助後續分析。若需求的預測不夠準確，會導致過少或是過多的存貨，過少會使得病患無法及時獲得治療、緩解，過多則容易造成濫用的風險，因此需求預測是至關重要的。 然而大部分的鴉片類藥物需求卻是所謂的間歇性需求(intermittent demand)，也就是在需求時間序列中伴隨著大量的零值，本研究使用某連鎖藥局從2021年11月底至2023年6月底，為期約2年的處方箋資料，本研究在僅有歷史需求資料的情境下，進行鴉片類藥品之需求預測。本研究將各藥品劃分為不同的間歇程度，並使用間歇性需求預測的傳統方法，此外也採用不同特徵組合於機器學習及深度學習之模型，進行比較分析。實驗結果發現，對於需求頻繁但波動大的藥品，可優先考慮使用多重滑動窗口下的XGBoost模型；而對於高間歇性藥品，則可採用LSTM搭配兩階段架構，提升其預測穩定性。若決策核心為某段時間區間內的預測準確率，則以 TSB 表現為最佳。

Opioid medications are frequently used in the treatment of cancer and chronic diseases, as they can effectively alleviate pain. With the increasing number of cancer and chronic disease patients and the trend of earlier onset ages, the demand for opioids has also grown year by year. However, opioids carry risks of dependence, addiction, and overdose, and in some cases, may even be diverted for illicit purposes to synthesize substances such as heroin, thereby posing significant threats to society. For these reasons, accurate demand forecasting can serve as a baseline for identifying abnormal prescription behaviors and assist in subsequent analyses. Inaccurate forecasts could lead to inventory imbalances: insufficient supply would prevent patients from receiving timely treatment and relief, while excessive supply could increase the risk of misuse. Thus, demand forecasting is of critical importance. Most opioid demand, however, exhibits what is known as intermittent demand, characterized by a high proportion of zero values in the demand time series. This study employs prescription data collected from a U.S. pharmacy chain between late November 2021 and late June 2023, spanning approximately two years. Under the condition of having only historical demand data, this research aims to forecast opioid demand. The study classifies drugs based on different degrees of intermittency and applies traditional intermittent demand forecasting methods, as well as machine learning and deep learning models with various feature combinations, for comparative analysis. Experimental results reveal that for drugs with frequent but highly volatile demand, the XGBoost model with multiple sliding windows performs best; whereas for highly intermittent drugs, the LSTM model combined with a two-stage architecture provides greater predictive stability. Furthermore, when the decision-making objective emphasizes forecast accuracy within a specific time horizon, the TSB method demonstrates superior performance.