Sevakula Rahul Kumar, Bota Patrícia J, Kassab Mohamad B, Bollepalli Sandeep Chandra, Thambiraj Geerthy, Boyer Richard, Isselbacher Eric M, Armoundas Antonis A
Cardiovascular Research Center, Massachusetts General Hospital, Boston, MA USA.
Anesthesia Department, Massachusetts General Hospital, Boston, MA USA.
NPJ Cardiovasc Health. 2025;2(1):41. doi: 10.1038/s44325-025-00075-5. Epub 2025 Aug 1.
Software-based blood pressure (BP) measurement offers non-invasive, continuous, real-time monitoring compared to traditional methods. This study explores a non-invasive machine learning approach to estimate arterial BP from ECG and SpO signals, using intra-arterial catheter BP readings as ground truth. A random forest (RF) algorithm was trained on a large dataset (~30 M beats, ~400 patient days), using extracted signal morphological features and patient characteristics. The RF model achieved low mean absolute error (MAE) for systolic/diastolic BP (4.29 ± 5.00 mmHg/2.38 ± 3.25 mmHg), independent of gender and race. Personalized models further improved performance (MAE: 3.51 ± 4.24 mmHg/1.85 ± 2.60 mmHg). We assessed different ECG lead combinations for estimating BP and observed that two limb leads, or a precordial lead were sufficient for an estimation below 5 mmHg MAE. These findings highlight the potential of real-time, personalized BP monitoring for early detection of hypertension, enhancing healthcare accessibility through non-invasive, continuous monitoring.
与传统方法相比，基于软件的血压测量提供了无创、连续、实时的监测。本研究探索了一种无创机器学习方法，以利用动脉内导管血压读数作为参考标准，从心电图（ECG）和血氧饱和度（SpO）信号中估计动脉血压。使用提取的信号形态特征和患者特征，在一个大型数据集（约3000万个搏动，约400个患者日）上训练随机森林（RF）算法。RF模型在收缩压/舒张压方面实现了较低的平均绝对误差（MAE）（4.29±5.00 mmHg/2.38±3.25 mmHg），与性别和种族无关。个性化模型进一步提高了性能（MAE：3.51±4.24 mmHg/1.85±2.60 mmHg）。我们评估了用于估计血压的不同ECG导联组合，观察到两个肢体导联或一个胸前导联足以实现MAE低于5 mmHg的估计。这些发现突出了实时、个性化血压监测在高血压早期检测中的潜力，通过无创、连续监测提高了医疗可及性。
