運用層級貝氏定理，結合危險率與活躍指標建立病人回診行為預測模型-以高血壓病人為例

A Bayesian Model for Forecasting Patient’s Return Visit Behavior with Hazard Rate Function and Customer Activity Index – Taking Hypertensive Patients as an Example

姜至曜(Chih-Yao Chiang)

指導教授 ： 任立中

國立臺灣大學/管理學院/國際企業學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900973

摘要

高血壓常被稱為「沉默的殺手」，由於我們的身體自己會去習慣血壓高造成的症狀，所以往往患者對血壓問題並無自覺。以台灣地區來說，106年高血壓性疾病的全民健保就診人數為3,796,606人，就診率為每十萬人口16,118人，且隨著台灣邁入高齡化社會，人數還在不斷的攀升，自然有更多的病人會前往各大醫院與診所尋求協助。因此，醫事機構單位面對高血壓病患就診人數不斷成長的情況下，需要建立相對應之模型協助醫師掌握病患就診行為，幫助形塑其就診病患輪廓，不僅可以降低備藥之不確定性，亦可以透過行銷手段防範病患之不定期回診與用藥之風險，進行後續的行銷策略規劃。
本研究使用某醫療機構位於六都的病患就診醫療數據庫，其資料期間為2015年6月1日至2017年6月30日，期望以此建立診所高血壓病患之回診行為預測模型。此模型以層級貝氏統計為基礎，透過先驗以及純先驗分配的設定，利用馬可夫鏈蒙地卡羅方法模擬參數的後驗分配，以吉氏抽樣連續抽樣，進而收斂至目標分配，再根據其估計的參數，結合危險率函數與顧客活躍性指標，建立病人回診行為預測模型，以期幫助診所掌握其病患之回診行為，藉以更準確地進行每期診所進藥量之預測，並且能夠針對不同顧客進行行銷策略的擬定，達到個別行銷的目的。

關鍵字

高血壓 ； 層級貝氏理論 ； 馬可夫鏈蒙地卡羅估計法 ； 半常態分配 ； 一般化反迦瑪分配 ； 危險率函數 ； 顧客活躍指標

並列摘要

Hypertension, also known as ‘silent killer’, has no obvious signs to indicate that something is wrong for patients to be aware of the disease as human bodies themselves are getting used to the symptoms caused by hypertension. Take Taiwan as an example, the number of clinic consultation of National Health Insurance for hypertension is 3,796,606, and the consultation rate is 16,118 per 100,000. As Taiwan has entered into an aging society, and the number of consultation rate keeps climbing, the fact that more people going to hospitals and clinics seeking for medical service is under expectation. Facing this phenomenon, medical institutions need to establish models which helps doctors to obtain patients’ return visit behaviors. For the further promotion of marketing plans, combining models with marketing strategies will not only reduce the uncertainty of prescription preparation, but also prevent the health condition risk of patients with irregular return visit behaviors from happening.
This research is based on a medical institution’s database of patient consultation from six municipalities from June 1st, 2015 to June 30th, 2017 and we use it to build a model for forecasting patient’s return visit behavior. The model is based on Hierarchical Bayesian model. With the set of prior and pure prior distribution, Markov Chain Monte Carlo method will stimulate parameter’s posterior distribution, and the posterior distribution will converge to the target distribution with Gibbs sampling method. Through this process, we can get the estimation of the parameter. Then, we can apply the estimated parameter to hazard rate function, and combine the function with customer activity index to build a forecast model of patients’ return visit behavior. With this model, we will be able to predict accurately on institutions’ requirements of drugs per period. Furthermore, it helps the institutions to formulate the marketing strategies for the purpose of direct marketing.

並列關鍵字

Hypertension ； Hierarchical Bayesian model ； Markov Chain Monte Carlo ； Half-Normal Distribution ； Generalized Inverse Gamma Distribution ； Hazard Rate Function ； Customer Activity Index

參考文獻

1.衛生福利部統計處(2018)。民國106年全民健康保險醫療統計。檢自https://dep.mohw.gov.tw/DOS/np-4267-113.html (Jun.15, 2019)

2.衛生福利部統計處（2016）。民國104年國人死因統計結果。檢自https://www.mohw.gov.tw/cp-2630-18831-1.html (Jun.15, 2019)

3.衛生福利部統計處（2017）。民國105年國人死因統計結果。檢自https://www.mohw.gov.tw/cp-16-33598-1.html (Jun.15, 2019)

4.衛生福利部統計處(2018)。民國106年國人死因統計結果。檢自https://www.mohw.gov.tw/cp-3795-41794-1.html (Jun.15, 2019)

5.陳夏蓮、蔡仁貞(2012)。症狀經驗、病因解釋與治療遵從行為--以高血壓病人為例。護理雜誌，59(1)，5-10。