AI Model Predicts ICU mortality in Heart Failure Patients

In the Intensive Care Unit (ICU), patients with heart failure often experience more severe and complicated conditions. This condition is commonly linked to multiple comorbidities, such as hypertension, coronary artery disease, chronic obstructive pulmonary disease, and chronic kidney disease. Consequently, ICU physicians must handle a wide range of clinical data, including patients’ medical histories, lab reports, imaging studies, prescribed medications, and continuous monitoring of vital signs. The sheer volume and complexity of this data make it difficult to quickly identify essential clinical information, thereby heightening the risk of misdiagnosis or delays in initiating treatment.

To overcome this issue, the healthcare sector has increasingly turned to artificial intelligence (AI) and big data technologies to support ICU physicians. While several predictive tools have been created to assess mortality risk in heart failure patients, these systems often depend on a combination of numerous variables and complex scoring frameworks. This poses a challenge since critically ill patients may not be suitable for extensive testing, and many smaller hospitals lack the diagnostic resources to support such assessments. In contrast, blood tests are simpler to perform and provide a comprehensive overview of the patient’s overall health. Now, an AI-driven model has demonstrated the ability to accurately forecast mortality risk in ICU patients with heart failure using only blood test data and vital signs.

A research team at Lanzhou University (Gansu, China) developed and trained nine AI models using 5,383 patient cases from the eICU database, along with an additional 530 cases from the MIMIC-IV database for external validation. These cases involved patients diagnosed primarily with heart failure, and the dataset covered a range of variables including age, blood oxygen saturation, white and red blood cell counts, platelet levels, hemoglobin, electrolyte levels, lactate, glucose, and other physiological and biochemical markers obtained during ICU admission. To improve the models’ reliability and ensure their broader applicability, all data underwent thorough preprocessing and standardization before model training.

The team applied several machine learning algorithms and used cross-validation to test the performance of the models, with the F1-score as the main evaluation metric. Of the nine models developed, the Multi-Layer Perceptron (MLP) achieved the best results, attaining the highest recall score of 0.64, which means it was most effective in identifying true positive cases and reducing missed diagnoses. It also recorded the highest performance scores in both the Macro F1 (0.74) and Weighted F1 (0.88) categories, indicating a strong balance across multiple classification metrics. The strong F1 results of the MLP model underscore its potential for clinical use in managing critically ill heart failure patients.