Blood sample analysis is a cornerstone of daily medical practice and is commonly employed for the detection of a large range of diseases. The recent availability of large-scale medical datasets, such as MIMIC-III, makes it feasible to evaluate the efficacy of blood samples for predicting a wide spectrum of diseases using machine learning (ML). A central challenge in this context is data quality of patient health records. Often patient data is incomplete or not informative enough. Here we address these problems with modern ML approaches to impute missing values and to incorporate auxiliary information to improve predictive performance. Our proposed clinical decision support system (CDSS) achieves an average state-of-the-art ROC-AUC score of 82.2% across the 50 most prevalent diseases in the MIMIC-III dataset. In order to validate our findings we complement the MIMIC-III data with survey data collected from trained medical doctors, who rated the predictive capacity of blood samples. Their ratings were compared to ROC-AUC values of the ML models. While Spearman’s \(\rho \) indicates a moderate agreement ( \(\rho = 0.515\) ), there are some diseases, where high ROC-AUC co-occurs with low predictability ratings from the medical doctors, i.e. some diseases were identified which are not commonly known to be detectable via blood samples. Finally, we derive explanations for individual predictions and demonstrate their reliability and clinical relevance in evaluations by medical experts.