📄ArXiv AI•較早收集於 2h
機器學習 vs 統計預測兒童肥胖
💡Simple logistic reg rivals XGBoost/TabNet on obesity data—rethink ML complexity for tabular tasks.
⚡ 30-Second TL;DR
有什麼變化
分析2021年美國兒童健康全國調查的18,792名兒童。
為什麼重要
研究顯示，邏輯斯迴歸等簡單模型在人口健康資料上常匹敵複雜ML，強調資料公平勝於演算法複雜度。此挑戰表格任務過度依賴深度學習。
下一步行動
Benchmark logistic regression vs XGBoost on your tabular health datasets to validate simple baselines.
誰應關注：Researchers & Academics
關鍵要點
- •分析2021年美國兒童健康全國調查的18,792名兒童。
- •預測因子涵蓋飲食、活動、睡眠、家長壓力、SES及社區特徵。
- •模型包括邏輯斯迴歸、RF、GBM、XGBoost、LightGBM、MLP、TabNet。
- •AUC 0.66-0.79；邏輯斯迴歸平衡最佳，提升模型改善召回。
- •種族及貧困群體效能差距持續存在。
🧠 深度解析
背景與延伸：來自公開資料，非原文內容。引用 6 個來源。
🔑 增強重點摘要
- •Sex-stratified and combined ML models using EHR data from up to five clinical encounters predict BMI before age 4 with MAE of 0.98 and R² of 0.72, showing no significant sex differences[1].
- •LSTM models using BMI at ages 3, 5, 7, and 11 achieve over 90% accuracy in classifying obesity at age 14 after SMOTE balancing, with MLP reaching 96% accuracy[2][3].
- •Novel predictors like facial images and kindergarten BMI Z-scores with demographics yield up to 87-92% accuracy in forecasting obesity, emphasizing early BMI data importance[3].
- •ML models for infant rapid weight gain (RWG) by age 1 using prenatal/postnatal data from multiple cohorts enable early intervention with acceptable accuracy in primary care[5].
🛠️ 技術深入
- •Sex-stratified models used 80/20 train/validation split with 5-fold cross-validation, evaluating MAE and R²; combined model optimal at MAE=0.98 (SD=0.03), R²=0.72 after five encounters averaging age 10.1 months[1].
- •Time-series models (ARIMA, XGBoost, LSTM, RNN) for BMI at age 10 had MAE 1.4-1.7, R² 0.48-0.54; LSTM slightly better for overweight, improved with resampling for balance[2].
- •Hybrid DT-LR for obesity risk via feature selection/classification; RF/GBoost on 190 variables for ages 6-9; LightGBM achieved 99.19% accuracy/F1 on obesity classification with 10-fold CV[3][6].
🔮 前景展望AI analysis grounded in cited sources
ML models will integrate into primary care for infant RWG screening by 2027
Models using routine prenatal/postnatal data show feasible accuracy for early population-wide obesity risk assessment before age 1[5].
Early BMI history from 5 encounters will standardize predictions before age 4
Combined models achieve reliable MAE 0.98 and R² 0.72, identifying 24 key variables without further improvement beyond five visits[1].
⏳ 時間線
2020
Taghiyev et al. introduce hybrid DT-LR for obesity prediction via feature selection
2020
Singh et al. evaluate 7 ML algorithms with SMOTE, MLP best at 96% accuracy for age 14 obesity
2021
Cheng et al. test 11 classifiers achieving 70% max accuracy; Zare et al. LR/ANN at 87% using kindergarten BMI
2021
Marcos-Pasero et al. apply RF/GBoost to 190 variables for BMI forecast in ages 6-9
2022
Cheng et al. LSTM models find 5 visits sufficient for BMI prediction before age 4, MAE 0.98
2025
ML models developed for infant RWG risk by age 1 using 7 cohorts for primary care integration
📎 來源 (6)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ArXiv AI ↗
每週 AI 簡報
每週一封，可隨時退訂。