Systemic acidemia critically impairs cardiovascular function and is typically diagnosed through arterial blood gas analysis, which may delay timely intervention. This study aimed to develop a deep learning model using electrocardiogram (ECG) data for rapid, non-invasive diagnosis of systemic acidemia. We developed and validated deep learning models to detect systemic acidemia, categorized by severity (mild: pH < 7.35, moderate: pH < 7.30, severe: pH < 7.20), using 12-lead ECGs from the Medical Information Mart for Intensive Care (MIMIC)-IV database. The training, validation, and test datasets included 56,249, 1,429, and 1,446 ECG-pH samples, respectively. To investigate variations in the model's efficacy depending on the type of systemic acidemia, we carried out a subgroup analysis according to pCO2 and HCO3- levels. The models achieved high performance in detecting systemic acidemia, with the area under the receiver operating characteristic curves (AUCs) of 0.69, 0.73, and 0.82 for mild, moderate, and severe acidemia, respectively, in the testing cohort. No statistically significant differences in AUC were observed between the group with pCO₂ ≤ 45 mmHg and HCO₃⁻ < 22 mEq/L and the group with pCO₂ > 45 mmHg and HCO₃⁻ ≥ 22 mEq/L, with p-values (DeLong’s method) of 0.07, 0.07, and 0.25 for mild, moderate, and severe acidemia, respectively. The group predicted to be positive by the severe academia model showed a lower 30-day survival rate compared to the negative group (p<0.05), supporting the clinical efficacy. Deep learning offers a non-invasive method for diagnosing systemic acidemia in critically ill patients, enabling earlier intervention and improved patient care.