於醫學隨機對照試驗結構化摘要中偵測PICO成份:比較首句及全句組

Abstract

本研究基於senetence-level在PubMed結構化文獻摘要中作PICO成份檢測，針對PICO個別成份每個段落的第一個句子是否足夠用來訓練樸素貝葉斯分類器以及支持向量機，抑或是除了段落第一句之外，還需涵蓋段落中除第一句之外的剩餘句子。從醫學資料庫 Pubmed 隨機對照試驗(randomized controlled trials)文獻中，擷取段首符合P/I/O各類別預設標籤的19,854結構化摘要來作樸素貝葉斯分類器以及支持向量機的訓練集。然後經由ten-fold cross-validation對於PICO各類別分別比較由段落第一句訓練出來的分類器(CF)以及整個段落裡所有句子訓練出來的分類器(CA)。採用recall, precision和F-measures作為結果加以比較。結果顯示使用樸素貝葉斯分類器，對於Outcome類別 CF 和 CA 之間並不存在顯著差異(F-measure=0.731±0.009 vs. 0.738±0.010, p=0.123)。然而在recall方面，CA在Intervention類別上表現得更好0.752±0.012 vs. 0.620±0.007, p<0.001)，F-measures 0.728±0.006 vs. 0.662±0.007, p<0.001)。對於Patient/Problem類別而言，CF具有更高的precision 0.714±0.009 vs. 0.665±0.010, p<0.001), 但較低的 recall 0.766±0.013 vs. 0.811±0.012, p<0.001)。在sentence-level PICO 成份偵測而言，CF並不總是優於CA，CF 和 CA在檢測不同PICO成份的表現各不相同。當使用線性支持向量機分類器時，P/I/O各類結果和使用樸素貝葉斯分類器並不完全相同，在sentence-level PICO 成份偵測而言，CF也並不總是優於CA，CF和CA在檢測不同PICO成份的表現也有所不同。而當支持向量機分類器使用radial based function時，recall以及F-measure都偏低。

To identify of patient, intervention, comparison, and outcome (PICO) components in medical articles efficiently is helpful in evidence-based medicine. The purpose of this study is to clarify whether first sentences of these components are good enough to train naive Bayes classifiers for sentence-level PICO element detection. We extracted 19,854 structured abstracts of randomized controlled trials with any P/I/O label from PubMed for naive Bayes classifiers training. Performances of classifiers trained by first sentences of each section (CF) and those trained by all sentences (CA) were compared using all sentences by ten-fold cross-validation. The results measured by recall, precision, and F-measures show that there are no significant differences in performance between CF and CA for detection of O-element (F-measure}=0.731±0.009 vs. 0.738±0.010$, p=0.123). However, CA perform better for I-elements, in terms of recall (0.752±0.012 vs. 0.620±0.007, p<0.001) and F-measures (0.728±0.006 vs. 0.662±0.007, p<0.001). For P-elements, CF have higher precision (0.714±0.009 vs. 0.665±0.010, p<0.001), but lower recall (0.766±0.013 vs. 0.811±0.012, p<0.001). CF are not always better than CA in sentence-level PICO element detection. Their performance varies in detecting different elements. In the study, support vector machines are also used. When comparing the results of CA and CF classifiers trained by Naive Bayesian and support vector machines, differences are obtained.