JMIR Publications

ABSTRACT

Background:

Long-term oxygen therapy (LTOT) improves outcomes in selected patients with severe chronic hypoxemia, but conventional LTOT uses fixed oxygen flow prescriptions that may not reflect changing needs during activity, sleep, or exacerbations. Artificial intelligence (AI) and automated oxygen systems may support continuous SpO₂ monitoring, signal-quality assessment, and adaptive oxygen titration. Evidence comparing AI-driven and non-AI automated approaches across performance, clinical readiness, LTOT applicability, and equity remains limited.

Objective:

This systematic review synthesized peer-reviewed evidence on AI-driven and automated systems for continuous SpO₂ monitoring or oxygen titration relevant to adult LTOT, focusing on accuracy, motion robustness, clinical performance, demographic equity, and readiness for home or ambulatory deployment.

Methods:

This review followed PRISMA 2020 and PRISMA-S. PubMed, IEEE Xplore, Springer Link, ACM Digital Library, and supplementary MDPI publisher-level searches were searched for English-language peer-reviewed studies published during 2000 to January 2025. Eligible studies evaluated AI-driven or automated systems for continuous SpO₂ monitoring or oxygen titration in adult LTOT-relevant populations and addressed motion artifact, low-perfusion signal management, skin tone bias, or prolonged signal stability. Two reviewers independently screened and extracted data; disagreements were resolved with a third reviewer. Risk of bias was assessed using an adapted ROBINS-I framework. Heterogeneous designs and outcomes precluded meta-analysis, so findings were synthesized narratively following Popay et al.

Results:

Of 928 records (926 from databases/platforms and 2 from manual reference screening), 912 remained after deduplication, 61 full texts were assessed, and 8 studies were included. Five studies evaluated AI-based systems and 3 evaluated automated non-AI oxygen delivery. AI models reported SpO₂ estimation MAE as low as 0.57% and RMSE as low as 0.69%, but most were simulated, retrospective, or non-COPD-specific. Only Cabanas et al. reported skin tone-stratified bias analysis. Automated systems showed stronger clinical deployment evidence: O₂matic maintained target SpO₂ 85.1% of the time versus 46.6% with manual titration, while Cirio and Nava reported mean SpO₂ of 95% versus 93% manually. Risk of bias was moderate to serious, mainly due to participant selection, limited demographic reporting, and algorithmic transparency.

Conclusions:

AI-driven and automated LTOT-relevant systems address complementary gaps. AI approaches show promise for signal interpretation and personalized prediction, whereas rule-based automated systems have stronger near-term clinical evidence for oxygen titration. Evidence is limited by small study numbers, heterogeneous outcomes, limited COPD/home LTOT validation, and sparse equity reporting. Future work should prioritize longitudinal validation in diverse LTOT populations, prespecified equity outcomes, failure-mode reporting, and hybrid architectures combining AI signal intelligence with safety-critical automated control.