<p>Performance metrics of RF models.</p>

Background Climate variability is increasingly recognized as a driver of child undernutrition, yet the non-linear relationships between specific climatic variables and nutrition remain unclear. This study uses machine learning to identify and quantify key climatic predictors of undernutrition among children. Methods In a mixed-method approach, a cross-sectional study assessed nutrition and child health outcomes in May 2024, while retrospective climate data was assessed spanning January 2022 to December 2023. The cross-sectional study recruited two hundred and seventy (270) children aged 6–23 months from rural areas in the Bosomtwe district. Anthropometry, hemoglobin concentrations, and food frequency were assessed using standard procedures. The collected data was standardized and subjected to principal component analyses to identify dietary patterns. Household food security was assessed using the USDA Household Food Security questionnaire, while the climate data was obtained from ERA5 reanalysis. Random forest algorithms were employed to evaluate the relative importance of various climatic factors in predicting undernutrition and morbidity. Decision trees were then derived from the models to examine interactions and thresholds. Results Rainfall emerged as the most critical climate predictors of stunting, while severe acute malnutrition was more sensitive to shortwave radiations. Temperature was the top predictor of fever, anaemia and diarrhoea. Low rainfall and high temperature substantially increased the possibility of undernutrition and morbidity. Threshold effects showed that rainfall below 4.78 mm and temperature under 23.40°C, increased stunting risk, especially when SW radiation drops below 6.01 W/m2. For severe acute malnutrition, rainfall below 5.53 mm and temperatures above 27.67°C significantly increased the risk. Conclusion This study finds significant influences of climate variables on child undernutrition, highlighting the importance of integrated climate health strategies that account for compound climate effects. These findings can inform the development of early warning systems and targeted interventions to mitigate climate-related health risks in vulnerable populations.背景 气候变率日益被认为是儿童营养不良的驱动因素，但特定气候变量与营养状况之间的非线性关联仍不明确。本研究借助机器学习技术，识别并量化影响儿童营养不良的关键气候预测因子。 方法 本研究采用混合研究设计，于2024年5月开展横断面调查以评估儿童营养与健康结局，同时回溯获取2022年1月至2023年12月的气候数据。本次横断面研究从博索姆特韦区的农村地区招募了270名年龄在6~23个月的儿童。采用标准流程开展人体测量学(Anthropometry)检测、血红蛋白浓度测定及食物频率评估。对收集到的数据进行标准化处理后，通过主成分分析识别膳食模式。采用美国农业部(USDA)家庭粮食安全问卷评估家庭粮食安全状况，气候数据则取自ERA5再分析资料。本研究运用随机森林(Random Forest)算法评估各类气候因子在预测儿童营养不良与发病情况中的相对重要性，并基于模型推导决策树以探究变量间的交互作用与阈值效应。 结果 降雨成为影响儿童生长迟缓(Stunting)的最关键气候预测因子，而严重急性营养不良(severe acute malnutrition)对短波辐射(shortwave radiations)更为敏感。温度则是预测发热、贫血与腹泻的首要因子。低降雨与高温会显著提升儿童营养不良及发病风险。阈值效应分析显示，当日降雨量低于4.78毫米、气温低于23.40℃时，儿童生长迟缓风险升高，尤其当短波辐射(SW radiation)低于6.01 W/m²时该效应更为显著。针对严重急性营养不良，降雨量低于5.53毫米且气温高于27.67℃会显著提升其发病风险。 结论 本研究证实气候变量对儿童营养不良具有显著影响，凸显了考虑复合气候效应的整合性气候健康策略的重要性。本研究结果可为早期预警系统的开发以及针对性干预措施的制定提供参考，以缓解脆弱人群面临的气候相关健康风险。