de Wijs C J, Behr J R, Streng L W J M, van der Graaf M E, Harms F A, Mik E G
Department of Anesthesiology, Erasmus Medical Center, Rotterdam, the Netherlands.
Faculty of Mechanical Engineering, Delft University of Technology, Delft, the Netherlands.
J Clin Monit Comput. 2025 Mar 30. doi: 10.1007/s10877-025-01291-1.
Monitoring in vivo mitochondrial oxygen tension (mitoPO) enables the measurement of mitochondrial oxygen consumption (mitoVO), providing deeper insights into the skin's mitochondrial environment. However, current mitoVO analysis often relies on manual identification of start and end points, which introduces substantial inter-user variability. Addressing this limitation is crucial for broader adoption, comparability, and reproducibility across research groups. Therefore, the aim of this study was to develop a neural network-based software that automatically analyzes mitoVO. A Bi-directional Long Short-Term Memory neural network was trained on 125 mitoPO measurement sequences and optimized through Bayesian optimization. It identifies start points and measurement periods, then applies a modified Michaelis-Menten fit to calculate mitoVO. This framework, embedded in automated software, was validated against the consensus of 3 raters. Bayesian optimization yielded an overall network performance of 94.2% on the test set. The neural network identified 91% of mitoVO start points within a ± 5-sample range of the manual consensus. Mean mitoVO values for the consensus and software were 6.56 and 6.63 mmHg s, respectively, corresponding to a bias of -0.057 mmHg s. Multiple runs of the network on the same dataset produced identical results, confirming consistency and eliminating inter-user variability. The developed neural network-based software automatically and consistently analyzes mitoVO measurements, substantially reducing reliance on subjective judgments. By enabling a standardized approach to mitoVO analysis, this tool improves data comparability and reproducibility across research settings. Future work will focus on further refining precision and extending functionality through multi-center collaborations.
监测体内线粒体氧张力（mitoPO）能够测量线粒体氧消耗（mitoVO），从而更深入地了解皮肤的线粒体环境。然而，目前的mitoVO分析通常依赖于手动识别起点和终点，这会引入显著的用户间差异。解决这一局限性对于各研究团队更广泛地采用、可比性和可重复性至关重要。因此，本研究的目的是开发一种基于神经网络的软件，用于自动分析mitoVO。使用125个mitoPO测量序列对双向长短期记忆神经网络进行训练，并通过贝叶斯优化进行优化。它可以识别起点和测量周期，然后应用修正的米氏方程拟合来计算mitoVO。这个嵌入自动化软件的框架根据3名评分者的共识进行了验证。贝叶斯优化在测试集上的整体网络性能为94.2%。神经网络在手动共识的±5个样本范围内识别出91%的mitoVO起点。共识和软件的平均mitoVO值分别为6.56和6.63 mmHg s，偏差为-0.057 mmHg s。在同一数据集上多次运行网络产生了相同的结果，证实了一致性并消除了用户间差异。所开发的基于神经网络的软件能够自动且一致地分析mitoVO测量值，大幅减少了对主观判断的依赖。通过实现mitoVO分析的标准化方法，该工具提高了不同研究环境下数据的可比性和可重复性。未来的工作将集中在通过多中心合作进一步提高精度和扩展功能。
