發展思覺失調症患者篩檢暨評估之多向度臉部情緒辨識測驗

Abstract

背景：臉部情緒辨識 (facial emotion recognition, FER) 是人們藉由觀察他人臉部表情以推論其情緒狀態之能力，通常包含7種情緒（快樂、悲傷、生氣、厭惡、害怕、驚訝及平靜）。思覺失調症患者常有中至重度之FER缺損，影響其精神症狀、社會功能及生活品質。然而，常用FER評估工具有4大缺失（內容向度不完整、計分指標不詳盡、未校正受試者性別之影響及心理計量特性大多未知），嚴重限制結果之解讀。此外，FER向度繁多，且信度及效率不易兼顧，故內容完整且精準之FER測驗，很可能題數過多，影響可行性。結合篩檢與詳細評估之測驗方式，可先快速辨識功能缺損之向度，再準確估計缺損嚴重度，或可改善前述信度與效率難以兼具之瓶頸。 目的：發展適用於思覺失調症患者篩檢暨評估之多向度臉部情緒辨識測驗 (Screening and Measuring Test of Multidimensional FER, SMART-FER)，並驗證其再測信度、練習效應、建構效度及已知族群效度。 方法：分為二個階段：（一）以3步驟發展SMART-FER：(1) 發展題庫：先自「專業表演者臉部表情常模資料」挑選7種情緒之照片做為候選題，再施測於患者及健康成人。研究者於剔除multidimensional Rasch model適配度不佳之題目，並考量性別differential item functioning (DIF) 後，剩餘題目組成最終版題庫。(2) 結合篩檢與詳細評估：藉由結合效能較佳之篩檢 (computerized classification testing, CCT) 及詳細評估測驗方法 (computerized adaptive testing, CAT），並比較不同終止條件下之篩檢效能（如正確率）、信度及施測效率，以挑選兼具高篩檢效能、高信度及高效率之組合，做為最終版SMART-FER；(3) 建構SMART-FER之施測介面。（二）驗證SMART-FER之心理計量特性：使用參與前一階段，以及願意接受再測（間隔4週後）且症狀穩定之患者資料，模擬分析SMART-FER之再測信度、練習效應及已知族群效度。 研究結果：第一階段共選168題候選題（7種情緒，各24題），並施測於351位患者及101位健康成人。於刪除3題適配度不佳之題目後，剩餘165題適配度良好 (infit and outfit mean square = 0.13–1.36)，支持其個別單向度，故納入最終版題庫。其中39題具嚴重性別DIF，故藉由因應受試者性別採用不同題目難度之方式，以校正其影響。由於高篩檢效能、高信度及高施測效率難以兼具，研究者改在特定篩檢效能下，挑選二種次佳之終止條件（快速模式及精準模式），以滿足使用者之需求。快速模式（「各向度篩檢7題」，搭配「信度 ≥ 0.70」或「多施測一題之信度增加量 < 0.001」之終止條件）之SMART-FER僅需68題（預計10分鐘），即具備可接受之篩檢效能（正確率 = 85.5%）及信度 (0.68–0.74)。精準模式（「各向度篩檢13題」，搭配「信度 ≥ 0.90」或「多施測一題之信度增加量 < 0.005」之終止條件），平均施測110題（約17分鐘），可達成良好之篩檢效能（正確率 = 91.8%）以及與完整題庫相似之信度（0.70–0.84 vs. 0.72–0.88）。共82位症狀穩定之患者參與再測評估。二種施測模式之SMART-FER皆具可接受至良好之再測信度 (intraclass correlation coefficient = 0.63–0.75 and 0.66–0.81）、已知族群效度（Cohen’s d = -0.48至-1.51及-0.49至-1.59）、建構效度（測驗結果與完整題庫高度相關，Pearson's r = 0.91至0.97），與微小至可忽略之練習效應（Cohen’s d = -0.15至0.23及-0.20至0.21)。 結論：初步結果顯示SMART-FER可提供完整（7種情緒且有個別向度分數）、有效（符合多向度模型，並能區分患者與健康人FER差異）且不受性別DIF影響之評估。此外，SMART-FER可彈性調整施測模式，分別強化測驗信度（精準模式）或效率（快速模式），以滿足不同使用者之需求。因此，SMART-FER具潛力廣泛應用於臨床及研究場域，以提升評估效能。

Background: Facial emotion recognition (FER) is the ability to identify others’ emotion status through their facial expressions, which contain identification of the 7 emotions (happiness, sadness, anger, disgust, fear, surprise, and calm). Patients with schizophrenia tend to have moderate to severe deficits of FER that affect their psychotic symptoms, social function, and quality of life. However, the commonly used FER measures have 4 flaws (i.e., incomprehensiveness, lack of score for each domain, possible biases due to examinees’ sex, and unknown psychometric properties), which severely limits their utility. Given the numerous domains of FER and the challenge of achieving high reliability and efficiency simultaneously, combining screening and measuring tests appears a promising solution. Purposes: The purpose of this study was to develop a screening and measuring test of multidimensional FER (SMART-FER). We examined its test-retest reliability, practice effect, ceiling and floor effects, construct validity, and known-groups validity. Methods: This study contained two phases. First, the SMART-FER was developed through 3 steps. (1) Forming the FER item bank. We first selected the candidate items (i.e., pictures of professional performers’ facial expressions across 7 emotions) from a database and validated them on patients with schizophrenia and healthy adults. The misfit items to the multidimensional Rasch model were removed. The items with differential item functioning (DIF) of sex were examined and considered. After that, the remaining items were used to form the item bank. (2) Combining screening and measuring tests: We first incorporated the two advanced testings: the computerized classification testing (CCT) and computerized adaptive testing (CAT). Then, simulations were performed to compare the accuracy, reliability, and efficiency of both tests with different combinations of stopping rules. We deemed the SMART-FER to be the test that achieved high accuracy, reliability, and efficiency simultaneously. (3) Constructing the administration system of the SMART-FER. Second, we examined the test-retest reliability and known-groups validity of the SMART-FER in patients with schizophrenia who had stable clinical severities and completed the FER item bank twice with a 4-week interval. Results: In phase 1, we selected 168 items (24 for each domain) as candidate items and tested these items on 351 patients with schizophrenia and 101 healthy adults. After removing 3 misfit items and adjusting item difficulties for the 39 DIF items, a total of 165 items were included in the FER item bank. All items showed good model fits (infit and outfit mean square = 0.13 to 1.36), supporting the unidimensionality of each domain. Given that high accuracy, reliability and efficient could not be achieved simultaneously, two alternative sets of rules (the “most reliable set” and the “most efficient set”) with acceptable accuracy were determined for prospective users. With the most efficient set (screening 7 items for each domain plus CAT with “reliability ≥ 0.70” or “limited reliability increase [LRI] < 0.001”), the SMART-FER needed 65 items (taking about 10 minutes) to achieve acceptable reliability (0.68–0.74) and accuracy (85.5%). Using the most reliable set (screening 17 items for each domain plus CAT with “reliability ≥ 0.90” or “LRI < 0.005”, the SMART-FER adopted about 110 items (17 minutes) to provide high accuracy (92.8%) and similar reliabilities to the FER item bank (0.70–0.84 vs. 0.72–0.88). In phase 2, 82 patients with stable symptom severities who completed the FER item bank twice. In general, with both assessment modes, the SMART-FER showed acceptable to good test-retest reliability (intraclass correlation coefficient = 0.63–0.75 and 0.66–0.81), trivial practice effect (Cohen’s d = -0.15 to 0.23 and -0.20 to 0.21), good construct validity (Pearson’s r with the FER item bank = 0.91 to 0.99), and satisfactory known-groups validity (Cohen’s d = -0.48 to -1.51 and -0.49 to -1.59). Conclusions: Our findings suggest that the SMART-FER provides comprehensive, valid, and unbiased assessments of patients’ FER levels. In addition, the stopping rules of the SMART-FER can be flexibly adapted to optimize the reliability or efficiency of assessments depending on users’ needs. Thus, it shows great potential to be applied in both clinical and research settings to improve the efficacy of assessments.