Magnetocardiography combined with machine learning for pulmonary hypertension detection and improved short-term risk assessment

Accurate detection and risk assessment are crucial for the prognosis of patients with pulmonary hypertension (PH). This study aimed to develop a detection model by employing ultra-high-sensitivity magnetocardiography (MCG) combined with machine learning while exploring the potential of MCG features in enhancing prognostic evaluation. PH patients were allocated into the exploratory and external validation cohort according to enrollment period in a single-center. The control group comprises patients with symptoms including dyspnea, chest tightness, chest pain, or fatigue in the absence of PH. Outpatient follow-up is performed for PH patients in the exploratory cohort. Seven machine learning algorithms and 15 out of 40 MCG features were employed to develop PH detection models. Cox regression is used to construct a risk assessment model. The exploratory cohort includes 312 subjects (156 PH patients), whereas the external validation cohort includes 162 subjects (81 PH patients). The Random Forest algorithm demonstrates superior performance, achieving sensitivities of 79.6% and 84.0% and specificities of 92.4% and 91.4% in different datasets, respectively. During the average follow-up period of 33 weeks, 14 patients experienced clinical deterioration events. Compared to the baseline model developed using three parameters from the guideline-recommended model, the combined model incorporating 2 MCG features performs better in this small cohort to detect deterioration events. In conclusion, the Random Forest-based MCG model demonstrated robust detection accuracy for PH. The integration of 2 MCG features with clinical baseline parameters may improve short-term risk assessment.