Focusing on chronic respiratory diseases, chronic obstructive pulmonary disease (COPD) is the major contributor to the global burden, with over 3.23 million deaths recorded in 2019, with 80% of these deaths occurring in low- and middle-income countries ^[1][2]. The average prevalence of COPD, as reported by European countries to the World Health Organization (WHO) from 2017 to the present, is between 1 and 2% and is expected to increase significantly by 2050 ^[2][3]. For example, it is forecasted that the prevalence of it will reach up to 13.8%, with a total of 4.8 million patients in Spain and up to 13.1% in United Kingdom. These predictions indicate a concerning trend of COPD becoming a major public health issue in the coming decades, highlighting the importance of continued research and development of effective interventions and treatments for this condition.

Long-term oxygen therapy (LTOT) is a common and well accepted treatment for patients with COPD and, in general, for patients with any other respiratory disease ^[3][4][5][4,5,6] tha strongly impacts patients’ capacity to breath ^[6][7][7,8]. Consequently, the main rationale behind LTOT is to improve this patients’ capacity to breath, thereby reducing breathing difficulties. To this end, patients are demanded to carry an oxygen concentrator, which is responsible for providing patients with the theoretically needed supplementary amount of oxygen. In order to measure these breathing difficulties, clinicians consider blood capillary peripheral oxygen saturation (SpO₂), which is the indicator used to assess the need for oxygen and conceptually stands for the fraction of oxygen-saturated hemoglobin relative to the total hemoglobin ^[7][8]. The supplementary oxygen flow rate prescribed to patients by clinicians is fixed and static and therefore, with a high level of probability, will not always match real patients’ oxygen needs ^[8][9]. This static dose is decided by clinicians applying a rule-based approach ^[7][8] that is based on the SpO₂ values obtained from patients running some well-known exercise tests (the most usual being the 6-min walking test, 6MWT). These tests are conducted periodically (e.g., twice per year) by patients in hospitals to adjust or modify their oxygen prescription based on the updated state of the disease ^[7][8].

However, the mismatch between the real oxygen needs and the oxygen provided often leads to either hypoxemia during exertion or hyperoxemia when resting. Put briefly, hypoxemia can cause adverse health effects such as dyspnea, muscle fatigue, and decreased exercise tolerance ^[9][10]. On the other hand, hyperoxemia, or an excess of oxygen in the blood, can cause health problems such as oxygen toxicity, which can damage the lungs and other organs ^[10][11]. Therefore, it would be extremely important to monitor in real time the SpO₂ levels in patients receiving LTOT and to dynamically adapt the oxygen flow rate to the real needs of patients in order to ensure that they always receive the appropriate amount of oxygen to avoid the problems mentioned above (see ^[11][12] for a more detailed analysis).

To address these challenges, the first approach might aim at designing an intelligent oxygen dosing system to dynamically and reactively deliver the appropriate dose of oxygen at the right time. In Researchersthis paper, we go far beyond this first approach and propose to to consider a reactive approach but rather to consider a proactive one that intends to personalize the treatment for each individual patient based on their specific physical conditions and the physical activity to be undertaken ^[12][13].

The reactive approach envisioned above is not new. Indeed, recent hospital tests have employed closed-loop oxygen dosing systems with good results ^[13][14]. These systems use control strategies based on proportional–integral–derivative, proportional–integral, or rule-based approaches to regulate the oxygen supply to a predefined set point. The oxygen dose is adjusted as a direct function of SpO₂, with satisfactory short-term results ^[14][15]. However, as expected, changes in oxygen saturation levels are not instantaneous and it takes several seconds for the delivered oxygen dose to affect the SpO₂ level ^[15][16], thus resulting in patients continuing to experience shortness of breath. Additionally, several tests that reswearchers have also conducted emphasized some fast and unexpected changes in the observed SpO₂ curve, which may turn into errors in the closed-loop decision process, thereby strongly affecting patients’ wellbeing. Aligned to theour view on the need for a personalized dosing strategy, Kofod et al. in ^[16][17] conducted a study that demonstrated the benefits of dynamic oxygen dosing. The authors concluded that “individualized automated oxygen titration decreases dyspnea during walking and increases walking endurance in patients with COPD on LTOT”.

The proactive approach proposed in this paper may efficiently address current limitations in the oxygen dosing system. To this end, by leveraging current tools and techniques available (e.g., edge–cloud continuum and artificial intelligence, AI) the strategy proposed in this paper will take advantage of the data generated by patients to create a personalized oxygen dosing system that will proactively adjust the dose to each patient’s specific situation. The system can predict future changes in SpO₂ levels based on past experiences and properly adjust the oxygen dose instantaneously to prevent heavy drops or increases in SpO₂. This adaptive approach can notably revolutionize the management of oxygen therapy for patients with chronic respiratory diseases, significantly improving patient wellbeing and their quality of life while reducing the burden of disease, extending the life of concentrators and their components, and ultimately benefiting the healthcare system. It is also critical to mention that the proposed proactive approach may be deployed on hospital premises but is specifically designed for ambulatory use, thus clearly impacting on real daily patient activity.

2. Personalized Oxygen Dosing System

In recent years, predictive models have gained significant attention in the respiratory field, with numerous studies conducted to improve the accuracy of predictions. Early contributions in this field date back to 2013 and were written by H. Elmoaqet et al. in ^[17][18], and considerable progress has been made to enhance the predictive capabilities of these systems since then. H. Elmoaqet et al. worked on predictive modeling for many years by applying mathematical methods even though no significant progress was obtained ^[18][19][19,20]. Their last contribution in the area ^[20][21] introduced a k-step predictive model utilizing AI for the first time. The authors proposed a framework for multi-step-ahead predictions of critical levels in physiological signals and developed a new performance metric for validation purposes (e.g., prediction accuracy). Results demonstrated a remarkable improvement compared with standard autoregressive models. Another recent study (Sam Ghazal et al. ^[21][22]) proposed the use of different machine learning (ML) techniques to predict five SpO₂ levels toward a proactive mechanical ventilator setting adjustment in an intensive care unit (ICU). The authors used an artificial neural network (NN) based on the back-propagation method along with a Bootstrap aggregation of complex decision tree implementation as the classifier. While the average results appeared promising, precision for medium and high severity events, which are the most important ones, was low, thereby highlighting the need for having better data quality and quantity to achieve acceptable results from predictive approaches. It is also worth mentioning that whatever predictive technology may be applied, the IT infrastructure should be able to properly accommodate the specific set of computation and storage requirements. The fog-to-cloud paradigm ^[22][23] (today also referred to as edge-to-cloud or more generically as cloud continuum) was introduced as a baseline technology for predictive systems. Xavi Masip et al. applied this paradigm to the health field and more particularly to the LTOT arena by utilizing patients’ contexts, historical data, and biological signals to proactively predict their oxygen dose needs in an optimal infrastructure utilization paradigm, which was centered on computing the results through edge devices ^[23][24]. This approach alleviates network load and distributes computing resources in a more efficient manner, demonstrating the feasibility of running AI models in edge devices using currently available technologies. More recently, H. Pascual et al. ^[24][25] provided an initial analysis aimed at predicting blood oxygen saturation by utilizing patients’ vital signs. To this end, a small size test with four patients was considered. The main objective was to assess the applicability of the same proposed AI architecture, which was trained with the same algorithm across diverse individuals, and analyze the obtained results. After analyzing the collected data, the authors concluded that, as expected for the sake of optimality, the same AI model architecture is not universally applicable to all patients. In the commercial arena, some products in the oxygen flow rate regulation field may be found, such as the O₂ Flow Regulator by Dima Italia™ ^[25][26] and Sanso Via™ by Sanso Health ^[26][27]. However, although these solutions focus on regulating patients’ O₂ flows, both of them are completely reactive (closed-loop) and only available on hospital premises. Another device on the market is iGo₂ ^[27][28], an oxygen concentrator designed to slightly adjust the oxygen dose to the patient’s activity through a rule-based system monitoring the respiratory rate. There is limited research available specifically addressing the challenges of personalized modeling and edge-computed SpO₂ predictions in the respiratory disorders area.