Machine learning integration identifying an eight-gene diagnostic signature for acute mountain sickness

Dan Yang^1,2Xinyao Yin³Qian Li²Xin Wang⁴Junqiang Gou¹Mengmeng Liu¹Xinman Peng¹Zhuxing Xu⁵Xiao Yang²Wenyan Jia¹Haiwen Tang¹Qiuli Zhang¹Feng Yang^1*Xiaofeng Wang¹Rui Wang¹

• ¹General Hospital of Xinjiang Military Region, Ürümqi, China
• ²Xinjiang Medical University, Urumqi, China
• ³New York University Shanghai, Pudong, China
• ⁴The nineth medical center of PLA general hospital Gynaecology and Obstetrics, Beijing, China
• ⁵Center for Disease Control and Prevention of ministry security in Xinjiang Military Region, Urumqi, China

Background: Acute mountain sickness (AMS) is highly prevalent at high altitudes, with estimated incidence rates ranging from 25% to 90%. However, current AMS diagnosis primarily relies on self-reported questionnaires, highlighting the need for reliable biomarkers. Thus, we aimed to establish a diagnostic model for AMS. Methods: We applied scRNA-seq (n = 10) and bulk RNA-seq (n = 192) to identify AMS-associated genes. Then, we constructed AMS diagnostic model by machine learning. We also assessed the expression levels of AMS-related gene signatures using Quantitative PCR. Finally, we explored the mechanism of AMS-associated signatures by epigenetic analyses and KEGG pathway enrichment. Results: We analyzed cellular heterogeneity through scRNA-seq data, revealing significant enrichment of myeloid (MD) and platelet (PLT) cells during AMS progression. Subsequently, we identified 526 differentially expressed genes (DEGs) associated with the progression of AMS using pseudobulk differential expression analysis on the MD and PLT subsets between the AMS and control groups. We further screened for AMS-associated genes using bulk RNA-seq based differential analysis and WGNCA. Finally, we screened 12 AMS-related genes using scRNA-seq and bulk-RNA-seq data. These genes were utilized as features across 113 distinct combinations of machine learning models to develop an AMS diagnostic model. The model of Stepglm[both] + NaiveBayes (ATP6V0C, BCL2A1, CD52, CSTA, GZMA, HINT1, PFDN5, and RNF11) demonstrated optimal diagnostic accuracy. It obtained an AUC of 0.948 on the training cohort (n = 160) and maintained robust performance on external validation cohorts, with AUCs of 0.818 (GSE103940 = 22) and 0.760 (GSE75665 = 10). Using qPCR, we confirmed that the mRNA levels of the model genes were aligned with the transcriptome data (P < 0.05). Based on the epigenetic analyses, we found the AMS signatures might regulate by the histone and m6A methylation. Furthermore, pathway analysis revealed significant enrichment of these signature genes in immune-related signaling pathways and oxidative stress (adjusted P < 0.05). Conclusion: Using machine learning, we identified and validated a minimal blood biomarker signature for AMS diagnosis. This approach offered a practical approach for the early detection of AMS, especially in resource-limited populations residing in high-altitude regions.