嬰幼兒人工智慧動作評估初探

Abstract

背景與目的：早產兒為神經發展障礙的高危險族群，其中又以動作發展障礙最早與最常出現。近年來，數位技術機革新及機器學習演算法進步，使得人工智慧模型能夠用以追蹤人體骨骼架構，這項進步也成為動作追蹤及動作辨識模型的基石。本研究之目的為使用深度機器學習方法為足月兒和早產兒開發人工智慧初步的動作識別模型。方法：本研究總共收集95名嬰幼兒，其中包括39名足月兒和56名早產兒，在年齡4至18個月時實施阿爾伯塔嬰幼兒動作量表評估以紀錄其動作發展變化。評估過程中有5部錄影機以不同角度同時進行拍錄，並對所有嬰兒進行追蹤，記錄其能獨自行走的年齡。資料分析的第一步由物理治療師依照阿爾伯塔嬰幼兒動作量表標準觀看影片並標註動作分類結果；影片標註的標註者內信度由一位物理治療師進行重複標註，標註者間信度由3位治療師標記。動作識別模型資料和治療師標記結果之比對，使用準確度、召回率、精確率和Ｆ值進行驗證。結果: 本研究初步資料分析顯示，數據前處理步驟的阿爾伯塔嬰幼兒動作量表實際評分的施測者間信度（ICC=0.96−0.99）與影片動作標註之評估者內（ICC=0.99−1, κ=1）及評估者間信度（ICC=0.91−1, κ=0.62-0.63）都達到良好的程度。資料的後處理包括95名嬰幼兒所進行之156人次評估影片，以及所貢獻之12,212個動作資料。針對31個嬰兒動作項目之初步機器學習模型的整體辨識準確度為0.89、召回率為0.90、精確率為0.89和Ｆ值為0.89。結論：本研究所建立的數據前處理步驟呈現評估者執行量表評估和動作標記的良好信度，初步的嬰兒動作辨識模型則顯示31項達成可接受的動作辨識準確度。後續須收集更多的資料量以提供機器訓練及精進，俾幫助納入更多動作項目以及改善有些易混淆的項目之準確度。

Background and Purposes: Preterm infants are at high risk of neurodevelopmental disabilities, with motor impairment being the leading problem. Breakthroughs in digital technology and advances in machine learning algorithms have enabled the tracking of the skeleton of the human body that creates a precedent for motor recognition and classification models. This study was aimed to use a deep machine learning approach to develop an artificial intelligence (AI) algorithm for the primary action recognition model in term and preterm infants. Methods: This study recruited a total of 95 infants that consisted of 39 term infants and 56 preterm infants. The infants were prospectively assessed for their motor function using the Alberta Infant Motor Scale (AIMS) from 4 to 18 months of age with their motor performance recorded by five cameras. All the infants were followed up for their age of walking attainment. The movement videos were labeled according to the scoring criteria of the AIMS items by three physical therapists and the movement annotations were then used for data processing and model training. The intra-rater reliability of the labeling annotations was examined in one therapist and its inter-rater reliability was examined in three physical therapists. The primary action recognition model was developed for training and testing on the movement video records against the annotated standard using the accuracy, recall, precision, and F-score. Results: The inter-rater reliability of administering the AIMS in the laboratory (ICC=0.96-0.99) and the intra-rater reliability (ICC=0.99−1, κ=1) and inter-rater reliability (ICC=0.91-1, κ=0.62-0.63) of the labeling procedure were good. For the data processing, 95 infants have contributed a total of 156 trials of videos that yielded 12,212 videos of AIMS items. The overall recognition accuracy of the preliminary machine learning model for 31 AIMS action items which compared with the annotated standards was 0.89, recall 0.90, precision 0.89, and F-value 0.89. Conclusion: The data pre-processing showed reliable infant motor assessment and labeling annotation. The preliminary action recognition model showed acceptable level of accuracy in classifying 31 movement items. More data collection is undergoing to include more movement items for machine learning and to enhance recognition accuracy of those easily confused movement items.