台灣餐飲場所蟑螂孳生風險之評估與預測從環境監測到機器學習模型之應用

Abstract

蟑螂為餐飲場所常見之衛生害蟲，除造成食品安全疑慮外，亦是重要的病原攜帶者。在台灣，現行餐飲衛生管理多依循食品良好衛生規範準則(Good Hygienic Practice, GHP)，側重於靜態的環境清潔稽核，缺乏對病媒族群動態的長期數據化評估。有鑑於此，本研究旨在利用病媒防治實務數據，透過縱向研究設計，釐清環境因子與蟑螂孳生間之關聯性，並結合機器學習演算法建立目擊風險預警模型與行動決策矩陣，以提供更精準的綜合性蟲害管理(Integrated Pest Management, IPM)策略。 本研究架構分為兩大區塊:第一部分與環夏台灣有限公司合作，利用其內部資料庫自2024年2月至2025年5月間，共177間餐飲場所的縱向追蹤資料，進行(1)蟑螂目擊事件與環境因子之關聯性分析，以及(2)目擊事件與生態結構之預測模型建立。第二部分則採前瞻性實驗調查設計，針對特定場域進行實地監測，旨在探討(3)室內微氣候對蟑螂生態之影響(2025年4月至2025年9月)，以及(4)硬體與衛生缺失對族群動態之影響(2025年3月至2025年9月)。統計分析採用廣義線性混合模型(Generalized Linear Mixed Model, GLMM)處理重複測量數據。並於預測模型建構的環節，比較GLMM、Random Forest與XGBoost三種演算法在預測效能上的差異。 從生物監測數據分析顯示，相對於對照組迴轉壽司店，烘焙店(蟑螂密度Relative Risk(RR)=13.66，偵測站陽性率Odds Ratio(OR)=7.21)與火鍋店(蟑螂密度RR=9.84，偵測站陽性率OR=6.09)為明顯高蟑螂活動的場所，但其人員目擊通報率卻存在顯著低估，且蟑螂密度與偵測站陽性率對目擊風險存在拮抗交互作用，顯示部分場所在嚴重蟲害下發生通報疲勞現象。在微氣候方面，發現環境濕度具有時段性差異:打烊後的高濕度顯著助長蟑螂密度(RR = 1.12)與偵測站陽性率(OR = 1.06)，而營業期間的濕度則因反映人類行為干擾而呈現抑制效應。在環境缺失方面，蟑螂族群結構受前期滯後效應顯著支配，惟結構孔洞(偵測站陽性率OR=1.03)與內場地面衛生(偵測站陽性率OR=1.03)仍顯著增加了族群擴散風險。 在預測模型效能方面，針對目擊事件預測，在發生與否之二元判斷上，Random Forest具備最高敏感度(Sensitivity=0.781)，適合用於早期預警，XGBoost 則在精確度(Precision=0.578)上表現最佳，利於資源配置；而在目擊次數之量化預測上，三種模型之誤差表現(平均絕對誤差與均方根誤差)相當，皆適用於未來的目擊次數估計。另在生態結構預測(蟑螂密度與偵測站陽性率)上，GLMM展現了最低的預測誤差，確立其作為未監測場所推估潛在蟲害結構的統計工具。 本研究整合上述發現，開發了「蟑螂目擊事件預測系統」與「蟑螂生態結構預測系統」，並建構「蟑螂防治行動參考值矩陣」，依據統計分位數將場所中的蟑螂危害嚴重度分級，指引餐飲場所管理者針對高密度或高擴散風險採取差異化的IPM策略。本研究證實單純遵循GHP不足以完全控制蟑螂孳生，建議主管機關與業者應轉型為數據治理模式，將常態性生物監測、微氣候監測及建立病媒目擊記錄SOP納入管理指標，以有效降低食品交叉汙染之風險。

Cockroaches are common hygienic pests in food service establishments. Beyond posing food safety concerns, they are significant carriers of pathogens. Current hygiene management in Taiwan primarily follows the "Good Hygienic Practice (GHP)" guidelines, focusing on static environmental inspections but lacking long-term, data-driven assessments of vector population dynamics. Therefore, this study aims to utilize practical pest control data through a longitudinal design to clarify the associations between environmental factors and cockroach infestation. Furthermore, it integrates machine learning algorithms to establish risk warning models and decision matrices, providing precise strategies for Integrated Pest Management (IPM). The study framework is divided into two parts. The first part, conducted in collaboration with HYSIA Taiwan Ltd., utilized their internal database covering longitudinal data from 177 food service establishments between February 2024 and May 2025. This phase analyzed (1) the associations between cockroach sightings and environmental factors, and (2) the development of prediction models for sightings and ecological structures. The second part employed a prospective experimental survey design to conduct on-site monitoring in specific venues, aimed at investigating (3) the impact of indoor microclimate on cockroach ecology (April 2025 to September 2025), and (4) the impact of hardware and hygiene deficiencies on population dynamics (March 2025 to September 2025). Generalized Linear Mixed Models (GLMM) were employed to handle repeated measures, and the prediction performance of GLMM, Random Forest, and XGBoost algorithms was compared. Analysis of biological monitoring data indicated that, compared to the control group (conveyor belt sushi restaurants), bakeries (Density Relative Risk (RR)=13.66; Positive Rate Odds Ratio (OR)=7.21) and hot pot restaurants (Density RR=9.84; Positive Rate OR=6.09) were industries with significantly higher cockroach activity. However, staff sighting reports in these venues were significantly underestimated. An antagonistic interaction was found between cockroach density/positive rate and sighting risks, suggesting a phenomenon of "reporting fatigue" under severe infestation. Regarding microclimate, humidity showed time-dependent differences: high humidity after closing hours significantly promoted cockroach density (RR=1.12) and positive rate (OR=1.06), whereas humidity during business hours showed a suppression effect due to human disturbance. Regarding environmental deficiencies, cockroach population structure was significantly dominated by the lag effect from the previous period; however, structural holes (Positive Rate OR=1.03) and poor internal floor hygiene (Positive Rate OR=1.03) still significantly increased the risk of population spread. Regarding prediction model performance, for the binary prediction of sighting events, Random Forest demonstrated the highest sensitivity (0.781), making it suitable for early warning, while XGBoost achieved the best precision (0.578), optimizing resource allocation. For the quantitative prediction of sighting counts, all three models exhibited comparable error rates in terms of Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). Conversely, for ecological structure prediction (density and positive rate), GLMM achieved the lowest prediction error, establishing it as a reliable statistical tool for estimating potential infestation structures in unmonitored sites. Integrating these findings, this study developed a "Cockroach Sighting Prediction System" and an "Ecological Structure Prediction System." Additionally, a "Cockroach Control Action Matrix" was constructed based on statistical quartiles to classify infestation severity, guiding managers to adopt differentiated IPM strategies for high-density or high-distribution risks. This study confirms that adhering solely to GHP is insufficient for fully controlling cockroach infestations. It is recommended that authorities and operators transition to a data governance model, incorporating routine biological monitoring, microclimate monitoring, and standardized sighting record SOPs into management indicators to effectively reduce the risk of food cross-contamination.