研究發現 AI 模型可以辨識有自閉症風險的幼兒，準確率達 80%

最近的研究表明，人工智慧 (AI) 在協助識別有自閉症風險的幼兒方面具有潛力，對於兩歲以下兒童的準確率約為 80%。

Artificial intelligence (AI) has shown promise in aiding the identification of toddlers at risk of autism, with an accuracy rate of about 80% for children under two, according to recent research.

最近的研究表明，人工智慧 (AI) 在幫助識別有自閉症風險的幼兒方面顯示出良好的前景，對於兩歲以下兒童的準確率約為 80%。

A team of researchers from the Karolinska Institutet in Sweden developed a machine learning-based screening system. While the AI model cannot replace traditional diagnostic methods, it could help identify children early on who may need further clinical evaluation.

瑞典卡羅林斯卡學院的一組研究人員開發了一種基於機器學習的篩檢系統。雖然人工智慧模型無法取代傳統的診斷方法，但它可以幫助儘早識別可能需要進一步臨床評估的兒童。

“Using [the] AI model, it can be possible to use available information and identify individuals with an elevated likelihood for autism so that they can get earlier diagnosis and help,” said Dr. Kristiina Tammimies, a study co-author.

研究共同作者克里斯蒂娜·塔米米斯(Kristiina Tammimies) 博士表示：「使用人工智慧模型，可以利用現有資訊並識別自閉症可能性較高的個體，以便他們能夠得到更早的診斷和幫助。

However, she cautioned that the model should not be viewed as a standalone diagnostic tool, reiterating that the final diagnosis should be conducted through standard clinical methods.

然而，她警告說，該模型不應被視為獨立的診斷工具，並重申最終診斷應透過標準臨床方法進行。

The AI model was developed using data from the U.S.-based Spark study, which provided information on 15,330 children diagnosed with autism and an equal number of children without the condition.

該人工智慧模型是利用美國 Spark 研究的數據開發的，該研究提供了 15,330 名被診斷患有自閉症的兒童和同等數量的非自閉症兒童的資訊。

From medical and background questionnaires, the researchers selected 28 measures that could be easily obtained before children reach 24 months of age, such as age at first smile, eating behaviors and age at first construction of longer sentences.

研究人員從醫學和背景問卷中選擇了28項在兒童24個月大之前可以輕鬆獲得的指標，例如第一次微笑的年齡、飲食行為和第一次構造較長句子的年齡。

Using machine learning to analyze patterns in the data, the research team compared the identified patterns between autistic and non-autistic children to build four different models, selecting the most effective one for further testing.

研究團隊利用機器學習分析數據中的模式，比較了自閉症和非自閉症兒童之間已識別的模式，建立了四種不同的模型，選擇最有效的模型進行進一步測試。

When applied to a separate dataset of 11,936 participants, the model correctly identified 78.9% of the children as either autistic or non-autistic. Specifically, the accuracy was 78.5% for children aged up to two years, 84.2% for those aged two to four years and 79.2% for those aged four to ten years.

當應用於包含 11,936 名參與者的單獨資料集時，該模型正確識別出 78.9% 的兒童為自閉症或非自閉症。具體來說，兩歲以下兒童的準確率為78.5%，兩歲至四歲兒童的準確率為84.2%，四歲至十歲兒童的準確率為79.2%。

An additional test using a dataset of 2,854 autistic individuals resulted in a lower accuracy rate of 68%, which the researchers attributed to differences in the dataset, including some missing parameters.

使用 2,854 名自閉症患者的資料集進行的另一項測試得出的準確率較低為 68%，研究人員將其歸因於資料集的差異，包括一些缺失的參數。

The study identified several key measures that significantly influenced the AI model's prediction of autism, including problems with eating certain foods, the age at which a child first constructed longer sentences, the age at which a child achieved potty training and the age at which a child first smiled.

該研究確定了幾個對人工智慧模型對自閉症預測有顯著影響的關鍵指標，包括吃某些食物的問題、孩子第一次構造較長句子的年齡、孩子進行如廁訓練的年齡以及孩子開始進行如廁訓練的年齡。

These factors, according to the research team, played a crucial role in the model’s ability to differentiate between autistic and non-autistic children.

研究小組表示，這些因素在模型區分自閉症和非自閉症兒童的能力中發揮了至關重要的作用。

Further analysis revealed that the model tended to identify autism more accurately in individuals who exhibited more severe symptoms and broader developmental issues. This finding suggests that the model might be more effective at recognizing cases with more noticeable developmental challenges accompanying autism.

進一步的分析表明，該模型傾向於更準確地識別出表現出更嚴重症狀和更廣泛發展問題的個體的自閉症。這項發現表明，該模型可能更有效地識別自閉症伴隨的更明顯的發育挑戰的病例。

Despite the promising results, some experts expressed concerns about the model's ability to correctly identify non-autistic children. With an 80% accuracy rate, the model could potentially lead to overdiagnosis and unnecessary stress for families, as 20% of non-autistic children might be incorrectly flagged as possibly autistic.

儘管結果令人鼓舞，但一些專家對該模型正確識別非自閉症兒童的能力表示擔憂。該模型的準確率高達 80%，可能會導致過度診斷並給家庭帶來不必要的壓力，因為 20% 的非自閉症兒童可能被錯誤地標記為可能患有自閉症。

Professor Ginny Russell from the University of Exeter sounded a note of caution regarding the push for early diagnosis, especially in very young children.

埃克塞特大學的金妮·拉塞爾教授對推動早期診斷，特別是對於年幼的兒童，發出了警告。

“It can be hard to tell the difference between a toddler who has a severe impairment and one who is simply developing more slowly but will eventually ‘catch up.’ I would not recommend applying psychiatric labels to children under the age of two on the basis of a limited range of behavioral indicators, such as whether they eat certain foods,” Russell said.

「很難區分患有嚴重障礙的幼兒和只是發育較慢但最終會『趕上來』的幼兒。我不建議根據有限的行為指標（例如是否吃某些食物）對兩歲以下的兒童貼上精神病學標籤，」拉塞爾說。