Poster R10 - Faster Detection of Hypoxia In Pilots From Vital Signs Data Using Dense Neural Network

Introduction: Hypoxia and reduced cerebral oxygenation remain a significant risk to military aviation and can all interfere with pilot performance, causing blackouts and death. Hypoxia, the leading physiological stressor to a pilot, is caused by a reduced level of oxygen availability that reduces the peripheral blood oxygen content in the circulating blood. These physiological changes can lead to poor cognition (poor decision-making), fatigue, vision changes, headache, and poor motor performance - all essential components of aviation performance. Hypoxia, generally defined as peripheral oxygen saturation (SpO2) under 90%, can lead to diminished cerebral perfusion. Respiratory rate is a vital indicator of hypoxia, but its direct measurement can be challenging. Fortunately, other physiological parameters like heart rate, pulse rate, and blood oxygen saturation (SpO2) can be used to estimate respiratory rate. For example, recent research on student pilots correlated higher ambient cockpit temperatures to higher heart rate (HR) and greater oxygen consumption (reduced SpO2) [1]. Understanding and detecting the correlations between brain oxygenation and physiological response for each pilot requires a precision medicine approach for which Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL) can help detect key features unique to each pilot, with a goal of earlier detection to enhance mitigation strategies. This study aims to develop an ML-based regression model using these physiological parameters to predict the respiratory rate and detect hypoxia. Additionally, this model can reduce the need for additional technology to monitor respiration rates in pilots, making it a more efficient and streamlined solution.

Materials and

Methods: This study utilizes dense neural network-based regression models to predict respiratory rate from heart rate, pulse rate, SpO2, and age. The dataset [2] includes simultaneous recordings of photoplethysmography (PPG), and respiratory signals acquired using conventional impedance pneumography (IP), with a sampling frequency of 125 Hz, from a cohort of 53 adult patients. The regression models were developed with 0, 1, 2, and 3 hidden layers with varying numbers of neurons to understand both linear and nonlinear relationships between physiological parameters and respiratory rate. The stochastic gradient descent algorithm was used to find the best weights of the regression model along with the dropout and elastic-net regularization to reduce overfitting during the training process. The model performances were evaluated using mean squared error and R2 score. A range of hyperparameter tuning was applied with batch sizes, epochs, learning rate, size of the dropout layers, different kernel initialization, and regularization with about 100 models to select the best model with the highest R2 score. By leveraging these physiological parameters, the model demonstrated better prediction accuracy in estimating respiratory rate, paving the way for effective hypoxia detection and thereby enhancing flight safety. This study can be improved further by further hyperparameter tuning and then comparing it with other deep learning techniques, such as recurrent neural networks (RNN), support vector machines (SVM), etc.

Results, Conclusions, and Discussions: RESULTS AND

Discussion:
In this paper, a deep learning model consisting of three dense neural network layers demonstrates a better performance in estimating respiratory rates for a specific age group. The highest R2 score was obtained for a 3-layer dense neural network containing 512, 256, and 128 hidden layers, respectively with a random normal weight initialization and L2 regularization on the hidden layers. The best model used a dropout layer with size 0.05 after the first hidden layer with a learning rate of 0.01, batch size of 200 and ran for 1000 epochs. The model achieves a high R2 score of 83% as shown in Figure 1, indicating a strong correlation between predicted and true respiratory rates, underscoring the model's ability to capture complex patterns and trends in the data. Additionally, Figure 2 displays the error distribution of the respiratory rate prediction model, with the x-axis representing the actual respiratory rate values and the y-axis showing the difference between true and predicted values. The plot reveals a uniform scatter of points around the zero-error line, indicating a well-distributed and unbiased prediction error. The compact clustering of points around the zero-error line suggests a high accuracy of the model in predicting respiratory rates, with minimal deviation from true values. This output performance suggests that the model presented in this paper can accurately detect hypoxia, enabling timely interventions and critical for flight safety.

Conclusions:
Integrated into aircraft systems, the model can continuously monitor pilots' vital signs, providing real-time alerts and warnings in the event of abnormal respiratory patterns. This proactive approach can prevent hypoxia-related incidents, reducing the risk of cognitive impairment, judgment errors, and even loss of consciousness. Moreover, the model's ability to adapt to individual pilots' physiological characteristics and flight conditions can provide personalized warnings, further enhancing its effectiveness. By predicting hypoxia with high accuracy, this model can safeguard pilots' health and well-being, ensuring safe and successful flights. Future refinements to the model can potentially enhance its predictive capabilities by comparative study, leading to even more accurate estimations and improved hypoxia detection.

Acknowledgements (Optional): REFERENCE:
[1] Shaw DM, Cabre G, Gant N. Hypoxic Hypoxia and Brain Function in Military Aviation: Basic Physiology and Applied Perspectives. Front Physiol. 2021 May 17; 12:665821. doi: 10.3389/fphys.2021.665821. PMID: 34093227; PMCID: PMC8171399.
[2] M. A. F. Pimentel et al., "Toward a Robust Estimation of Respiratory Rate from Pulse Oximeters," in IEEE Transactions on Biomedical Engineering, vol. 64, no. 8, pp. 1914-1923, Aug. 2017, doi: 10.1109/TBME.2016.2613124.