請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89215
標題: | 人工智慧應用於嬰幼兒動作評估應用程式 Application of Artificial Intelligence in Mobile Application for Infant Movement Assessment |
作者: | 蕭郁靜 Yu-Ching Hsiao |
指導教授: | 鄭素芳 Suh-Fang Jeng |
關鍵字: | 早產兒,居家影片,家長辨認能力,機器學習,嬰幼兒動作發展評估, preterm infants,home videos,parental perception,machine learning,infant motor assessment, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 背景與目的:早產兒作為神經發展障礙的高危險族群,動作損傷是主要的問題之一,也因此追蹤早產兒的動作發展至關重要。家長作為第一線觀察嬰幼兒動作發展的觀察者,有機會藉由拍攝嬰幼兒居家活動影片提供嬰幼兒發展的資訊給專業人員評估。近年來隨著人工智慧的發展,該技術逐漸廣泛地使用在評估人類動作然而卻鮮少應用在嬰幼兒動作發展評估。本研究的目的是探討家長對於嬰幼兒動作的辨認能力以及結合手機應用程式建立動作辨識模型用以辨識足月兒及早產兒的動作。
方法:本研究招募月齡四個月的早產兒37位和足月兒54位並使用阿爾伯塔嬰幼兒動作評估量表追蹤受試者4-18個月月齡的動作發展。本研究團隊開發手機應用程式“Baby Go”以收集家長在家中拍攝的阿爾伯塔嬰幼兒動作評估量表中的動作影片,收集影片使用在模型訓練跟動作辨識。手機應用程式有教學指引以引導家長拍攝優良品質的居家影片。家長的嬰幼兒動作辨認能力會以家長選擇上傳的動作項目與物理治療師回顧居家影片後判斷動作項目一致的百分比呈現。居家影片上傳完成後,物理治療師會標註影片中的動作項目,再將標註結果交予合作的資訊工程學系研究團隊進行機器學習。物理治療師標註的動作項目作為動作辨識的黃金標準與動作辨識模型的動作辨識結果進行比較。 結果:家長可以準確地辨認嬰幼兒的躺姿及趴姿動作(一致性>65%)但較無法精準辨認嬰幼兒的坐姿及站姿動作。身體缺失跟不穩定的運鏡是影響居家影片品質的常見因素,其中身體缺失佔全部標註資料的27.7%,運鏡不穩定佔全部標註資料的27.4%。家長總共上傳1027個居家影片,動作辨識模型辨認31個阿爾伯塔嬰幼兒動作評估量表動作項目的準確率是0.77、精確率是0.66、召回率是0.66以及F值是0.65。 結論:家長可以精準辨認嬰幼兒躺姿跟趴姿動作,動作辨認模型辨認居家影片中的嬰幼兒動作呈現中度準確率。未來需要收集更多居家影片以及納入多樣化的樣本以提升動作辨識的準確率並增加動作辨識模型的使用普及性。 Background and Purposes: Preterm infants have increased risk of neurodevelopmental disabilities with motor impairment being the major problem. Parents are the first-line observers of their infant’s motor development that are likely to provide helpful developmental information through video recording at home. Artificial Intelligence (AI) approach is increasingly developed for human movement assessment, however, it’s application on infant motor assessment is rarely explored. This study aimed to examine parental perception of infant movements and developed action recognition model incorporation of an application (APP) to identify movements in preterm and term infants. Methods: This study included 37 preterm infants and 54 term infants, and prospectively followed up their motor development using the Alberta Infants Motor Scale (AIMS) (58 items) from 4 to 18 months of age. The APP “Baby Go” based on the AIMS items was developed for their parents to upload infant movements at home for subsequent machine learning and movement identification. The APP was equipped with an education module to guide parents regarding good-quality video recording. Parental perception of infant movements was examined using the agreement of parent’s classification of AIMS items compared with the physiotherapist’s results. Home video files were annotated by the physiotherapist and served as the standards to examine the accuracy of AI algorithm. Results: Parents were accurate in classifying supine and prone movements (agreement >65%) but were less likely to classify sitting and standing movements. Missing body parts and unstable camera movement were common factors affecting the quality of home videos with 27.7% of annotated results showed missing body parts and 27.4% of annotated results showed unstable camera movement. The parents uploaded 1,027 home videos via the APP and the overall accuracy of the action recognition model in classification of 31 AIMS items was 0.77, precision was 0.66, recall was 0.66, and F-score was 0.65. Conclusion: Parents showed accurate perception of infant movements particularly occurring in supine and prone position. The overall action recognition model of home videos achieved moderate level of accuracy. Future work needs to increase the number of home videos and the heterogeneity of sample to enhance the accuracy of action recognition model and generalizability of the results. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89215 |
DOI: | 10.6342/NTU202303122 |
全文授權: | 未授權 |
顯示於系所單位: | 物理治療學系所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 5.07 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。