Machine learning methods for detecting rare fungal and tuberculosis infections in patients with hepa
Recently, global health institutions and public health systems have continued to advance service capacity building to provide stronger guarantees for public health. Diagnosis of pathogen-induced infectious diseases is a complex and highly challenging problem given their unique characteristics. A-priori characteristic identification using modern AI methods can facilitate the diagnostic process, potentially having a positive impact on improving outcomes, early diagnosis and efficient resource use.
Our goal is to develop an AI-driven clinical decision support system for identifying rare and specific infections - including a complex composed of six diseases with significantly different microbiological and clinical specificities -- including osteosarcoma, hypoglycemia, tuberculosis, mesothelioma, pulmonary meningitis, pneumococcal meningitis, and tuberculosis -- unified into their shared diagnostic delay patterns - in the early stages of hospitalization, often due to delays caused by unspecified symptoms, predetermined biases, low prognosis, and clinical suspicions, leading to a uniform delay of 20-30 days. We analyzed data collected from 11,494 patients with rare infections, 10,000 controls at Mayo Clinic Hospital (2010-2023) during their stay. Based on data collected from patients at Mayo Clinic Hospital during their stay of three days, machine learning models, including HistGradientBoosting, were trained in the characteristics and categories of data collected from patients at Mayo Clinic Hospital during their stay of three days.
Class imbalances are treated, class imbalances are treated through SMOTE, and SHAP values provide explanation. Performance is assessed as accuracy, sensitivity, specificity, and area under the ROC curve using accuracy, sensitivity, specificity and area under the ROC curve as performance evaluation indicators. The HistGradientBoosting model has an accuracy of 93.
2%, an AUC of 0. 947 (IQR of 0. 942-0.
952)), significantness (97. 3%) and sensitivity (61. 9%) in the predictive value program under the default classification thresholds of the model.
Key predictive indicators include a decrease in high phosphoric acid levels, a decrease in lymphocytes counts, changes in male sex, signs of immune deficiency, age, and levels of phosphorus, which are consistent with clinical marker indicators and show strong potential for early diagnosis of rare infections. By using electronic health record data and advanced machine learning models, the system can improve resource efficiency, make clinical interventions timely, and ultimately improve patient effectiveness. External data and pioneering experiments further validate these findings and facilitate their implementation in the real world.
It should be noted that these are compiled based on international public research findings, for reference in HealthCop, and do not constitute clinical advice. If you have health questions, consult professional medical professionals.
