遠端嬰幼兒動作評估行動應用程式「寶貝動起來」應用於早產和足月兒的使用性

Abstract

背景與目的：早產兒通常伴隨較高的神經發展障礙風險，尤其在出生後第一年常見動作發展遲緩。因此，及早評估其動作發展並提供早期介入至關重要。隨著數位科技的進步，越來越多遠距發展評估工具應運而生，讓家長能透過手機在家中拍攝嬰兒的動作影片並上傳進行評估。人工智慧技術也逐漸被應用於嬰幼兒動作發展分析，但大多限於嬰兒早期的仰姿動作。本研究旨在開發一款結合人工智慧技術的手機應用程式 “Baby Go”，協助家長遠距評估嬰兒時期多姿勢下的動作發展，同時探討家長與臨床人員對此應用程式的使用情況與可行性。方法：本研究共招募32位早產兒與19位足月兒及其家長，在嬰兒2至18個月大之間使用手機應用程式 “Baby Go” 第二版評估動作發展。家長需進行嬰兒38項多姿勢之動作自評，並上傳影片以進行人工智慧模型分析，結果會在一至兩天後呈現。受訓過之物理治療師根據阿爾伯塔嬰兒動作量表的評估標準對影片進行標註，並評估影片品質，同時記錄家長使用手機應用程式的情形、使用過程中是否需要協助，以及人工智慧、家長與物理治療師三者間對嬰兒動作評估結果的一致性。家長會在嬰兒達6、12與18個月時填寫可用性問卷，物理治療師則在評估結束後填寫可用性問卷。研究結果：共有29位早產兒（93.5%）和16位足月兒（84.2%）至少使用應用程式上傳過一部影片，而92.3% 的影片品質為中等以上。家長上傳影片的數量隨嬰兒月齡增加而逐漸遞減，3到5個月的動作評估有最高的完整上傳率（50.0%到54.5%），然而超過50%的家長在嬰兒滿13個月後就不再上傳。排除直向拍攝或嬰兒方向錯誤的影片後，家長、人工智慧與物理治療師在動作評估結果的一致率分別為79%、79%與73%。大多數家長認為 “Baby Go” 使用起來簡單（93%），也幫助他們更了解嬰兒的發展狀況（100%）。但部分家長反映錄影的條件太嚴格及有時難以捕捉嬰兒的動作表現。物理治療師則認為 “Baby Go” 能夠有效地輔助篩檢嬰兒的動作發展。結論：家長使用行動應用程式“Baby Go”第二版上傳至少一段影片的上傳率高，多數家長也認為“Baby Go”行動應用程式容易操作。然而，人工智慧與物理治療師在嬰幼兒動作辨識上的一致率尚有不足，顯示仍有提升人工智慧模型準確度的空間。未來可能須調整部分動作項目，並持續增進應用程式的功能，以進一步提升其可用性與準確度。

Background and Purposes: Preterm birth is associated with a higher risk of neurodevelopmental disorders, with motor impairment as the major disorder during the first year of life. Early motor assessment is essential for identifying potential motor delays and enabling timely intervention in preterm infants. With the advancement of digital technologies, remote screening tools have been developed to assess infant motor development at home based on parent-recorded videos. Artificial intelligence (AI) has shown promising potential in supporting physiotherapists in evaluating infant motor development; however, the applications were mainly for supine movements in early infancy. This study aims to develop a mobile application (APP), “Baby Go”, that integrates AI technology for remote infant motor assessment in various positions throughout infancy, and to investigate its usability from the perspectives of both parents and clinicians. Methods: This study enrolled 32 preterm and 19 full-term infants and their parents using the “Baby Go” APP to assess their infants’ motor development from 2 to 18 months of age. Parents were asked to perform parental perception assessment of 38 movements, and upload videos of the observed movements for AI classification. After uploading, parents received developmental results generated by AI assessment. Physiotherapists annotated the uploaded video using the Alberta Infant Motor Scale (AIMS) criteria. The APP usage, external support for APP use, quality of videos, and agreement between the AI, parents’, and physiotherapists’ labeling results were examined. Parents provided feedback on the APP usability when their infants approached 6, 12, and 18 months of age. Two physiotherapists provided feedback on the APP usability when their patients reached 18 months of age. Results: Twenty-nine (93.5%) preterm and 16 (84.2%) full-term infants uploaded at least one video using the “Baby Go” APP version 2. Of the 936 uploaded videos, 864 (92.3%) showed fair to high quality. Uploaded frequency decreased gradually with age. Most parents completed video uploads during the 3- to 5-month assessments, with more than 50% failing to upload any videos beyond 13 months of age. After excluding those videos with incorrect recording format, the agreement between the AI, parents’, and physiotherapists’ results for each movement were 79%, 79%, and 73%. The usability survey from parents revealed that 93% considered the APP easy to use, and all parents reported that the APP helped them better understand infant development. However, some parents reported that the recording criterion is rigid and difficult to meet, and the baby’s movement is hard to capture. The clinicians indicated that the APP helps screen infant motor development. Conclusion: The upload rate by parents of preterm and full-term infants using the “Baby Go” version 2 was high, and most parents reported that the APP was easy to use. However, the moderate agreement between the AI and physiotherapists’ results highlights the need to improve the AI model’s accuracy. Future work is necessary to enhance the AI model accuracy and adjustment of movement items to further increase the APP’s usability.