The findings come from the multicenter SAVE-O2 AI trial, led by researchers at the University of Colorado Anschutz Medical Campus and published in JAMA Internal Medicine. The results were presented simultaneously at the Military Health System Research Symposium, highlighting the technology’s possible value not only in hospitals but also in military and emergency-care environments where clinical staff may be stretched thin.

The trial enrolled 300 adults at four U.S. hospitals, including UCHealth University of Colorado Hospital. Participants had acute respiratory illnesses, traumatic injuries, burns or conditions requiring surgical recovery, and all had recently begun receiving supplemental oxygen. They were randomly assigned either to conventional oxygen management, in which nurses or respiratory therapists adjusted flow rates, or to autonomous oxygen titration using the investigational O2matic PRO100 system.

Supplemental oxygen is usually delivered through devices such as nasal cannulas or face masks, with the flow rate adjusted according to intermittent measurements of a patient’s blood oxygen saturation. That saturation is estimated by pulse oximetry, a noninvasive technique that uses light to detect changes in the color of blood circulating through a fingertip sensor. Although pulse oximeters can provide continuous readings, standard hospital practice generally relies on clinicians checking the values periodically and manually changing oxygen delivery.

The automated system used the same basic physiological signal but responded to it continuously. When the patient’s oxygen saturation moved below the prescribed range, the device could increase oxygen flow; when the level rose too high, it could reduce delivery. This closed-loop approach is designed to compensate for the rapid fluctuations that can occur as patients breathe, move, sleep, receive medication or experience changes in their underlying illness. Instead of waiting for the next clinical assessment, the system adjusted oxygen in near real time.

Patients assigned to automated therapy spent 85 percent of the monitored period within their target oxygen range, a 22-percentage-point improvement over standard care. They also spent less time in hypoxemia, the condition in which blood oxygen levels fall too low, and less time in hyperoxemia, when oxygen levels exceed the intended range. The investigators further reported that clinical staff made fewer manual adjustments for patients in the automated group.

The distinction between too little and too much oxygen is clinically important. Insufficient oxygen can deprive organs such as the brain and heart of the oxygen they need to function. Excess oxygen, once widely assumed to be harmless, may also cause problems in some critically ill patients, including oxidative stress and injury to vulnerable tissues. For that reason, modern oxygen therapy increasingly emphasizes maintaining a patient-specific target range rather than simply delivering as much oxygen as possible.

“Oxygen is one of the most widely used therapies in medicine,” said Adit Ginde, the study’s principal investigator and a professor of emergency medicine at the University of Colorado Anschutz School of Medicine. Yet oxygen delivery remains largely dependent on repeated manual adjustments. According to Ginde, autonomous titration could help patients stay within their intended range more consistently while reducing both under-oxygenation and over-oxygenation.

David Douin, the study’s first author and an associate professor of anesthesiology at the University of Colorado Anschutz School of Medicine, said hospitalized patients’ oxygen requirements can change quickly. An automated system can react to those changes throughout the day and night, potentially reducing the periods during which a patient’s oxygen level drifts outside the target range. The trial found no increase in serious adverse events, an important safety result for a device that directly influences a core component of respiratory support.

The researchers caution that improved oxygen control does not by itself prove that the technology improves survival, shortens hospital stays or prevents long-term complications. The study primarily evaluated how much time patients spent within their prescribed oxygen range and how often staff needed to intervene. Future investigations will need to examine clinical outcomes, workload changes, performance in more severely ill patients and the system’s reliability during transport or in settings with limited monitoring resources.

The research has particular relevance to military medicine, where medics may care for wounded service members far from a hospital while simultaneously managing bleeding, airway problems and other life-threatening injuries. Vik Bebarta, chair of emergency medicine and founding director of the CU Anschutz Combat Medicine Research Center, said a device capable of adjusting oxygen independently could remove one recurring task from an overloaded medic’s responsibilities. The team is now planning additional evaluations in emergency transport and prehospital care.

The O2matic PRO100 was investigational in the United States and had not been cleared or approved by the Food and Drug Administration for commercial use. It was rented from O2matic of Denmark for research purposes, while the company had no role in study design, data collection, analysis or publication decisions. The trial was conducted under an FDA Investigational Device Exemption and was supported by the Defense Health Agency’s Combat Casualty Care Portfolio, the Medical Technology Enterprise Consortium and the National Center for Advancing Translational Sciences.

Subject of Research: Automated oxygen delivery and real-time oxygen titration for hospitalized patients.

Article Title: SAVE-O2 AI trial of autonomous oxygen titration in hospitalized adults.

Web References: JAMA Internal Medicine article; University of Colorado Anschutz; O2matic.

References: Ginde A, Douin D and colleagues, SAVE-O2 AI trial, JAMA Internal Medicine; CU Anschutz Combat Medicine Research Center.

Keywords: artificial intelligence, oxygen therapy, automated oxygen delivery, pulse oximetry, hypoxemia, hyperoxemia, hospital medicine, emergency medicine, military medicine, clinical trial.

This multicenter, retrospective study compiled data from 206 HCC patients with pathologically confirmed diagnoses, sourced from three distinct hospitals, ensuring a diverse and robust dataset for model training and validation. The investigative team focused sharply on the hepatobiliary phase images (HBP) of gadoxetic acid-enhanced MRI, a specialized imaging sequence known for its superior liver lesion characterization. To harness the full potential of this imaging modality, three variations of deep learning architectures were meticulously developed and assessed: two-dimensional (2D), three-dimensional (3D), and a novel 2.5-dimensional (2.5D) deep multi-instance learning (MIL) model.

Among these approaches, the 2.5D MIL technique emerged as the most potent, ingeniously integrating information from all axial slices encompassing the tumor and surrounding peritumoral regions. This comprehensive slice selection allowed the model to capture both intratumoral heterogeneity and critical peritumoral microenvironmental features, which are hypothesized to play pivotal roles in MVI development. Remarkably, this approach outperformed conventional models with area under the curve (AUC) values reaching 0.802 in internal validation and 0.759 in the external test cohort, highlighting its strong generalizability across patient populations.

Building on these promising findings, the researchers extended their methodology by incorporating additional MRI sequences—T1-weighted fat-suppressed (T1WI-FS) and T2-weighted fat-suppressed (T2WI-FS) images—into a multimodal prediction framework. The synergistic use of these complementary sequences aimed to further refine the predictive algorithm by capturing distinct tissue contrasts and pathological signatures of MVI. This multimodal deep learning model demonstrated exceptional predictive capacity, with AUC values soaring as high as 0.954 in the training set, and sustaining robust performance metrics with AUCs of 0.857 and 0.788 in independent validation and test sets respectively.

These advancements are not merely incremental; they represent a paradigm shift in medical imaging diagnostics for liver cancer. The exploitation of 2.5D MIL in this context signifies a novel computational strategy that leverages the spatial and contextual richness of MRI datasets more effectively than traditional machine learning or purely 2D/3D frameworks. By embracing the heterogeneous nature of tumor microenvironments across consecutive axial slices, this approach unlocks a more nuanced understanding of tumor biology, which is crucial for preoperative risk stratification.

Importantly, the ability to noninvasively predict MVI preoperatively holds profound clinical implications. Surgeons and oncologists can now potentially tailor treatment regimens with increased precision—deciding between surgical resection, transplantation, or adjuvant therapies based on individualized risk profiles. Moreover, early detection of MVI propensity may guide surveillance intensity post-operation, improving patient outcomes through timely interventions.

The retrospective design of this study, coupled with its multicenter data acquisition, lends credibility to the model’s robustness and applicability across different clinical settings. However, the authors underscore the necessity for prospective trials to validate and eventually integrate these computational tools into routine clinical workflows. Additionally, further research into integrating radiogenomics could uncover deeper mechanistic links between image features and genetic underpinnings of microvascular invasion.

On the technical front, the study exemplifies an exemplary application of advanced artificial intelligence methodologies in oncological imaging. The use of deep learning architectures capable of handling multi-instance learning tasks shows how algorithmic innovation can extract actionable insights from complex and high-dimensional medical images. Fine-tuning such models with diverse MRI sequences reinforces the importance of multimodality in capturing the multifaceted nature of cancer pathology.

This research also emphasizes the crucial role of peritumoral tissue analysis, an often-overlooked area in conventional imaging assessments. The findings suggest that changes in the tumor microenvironment surrounding hepatocellular carcinoma lesions carry significant predictive information, urging the clinical community to expand diagnostic focus beyond tumor margins. Such insights will likely spur new investigations into tumor-stroma interactions and their implications in cancer progression.

Future directions proposed by the research team involve the development of automated MRI preprocessing pipelines and user-friendly software interfaces to democratize the use of their predictive models in community hospitals and cancer centers worldwide. Integrating these outputs with electronic health records and decision-support systems could democratize access to personalized oncology care, aligning with global efforts toward precision medicine.

In conclusion, the creation and validation of a deep multi-instance learning model based on gadoxetic acid-enhanced MRI heralds a new age in the preoperative management of hepatocellular carcinoma. By harnessing the power of 2.5D imaging data and multimodal sequences, this approach offers an unprecedented window into tumor invasiveness, enabling clinicians to make better-informed decisions and ultimately improving patient prognoses. As artificial intelligence continues to evolve, such innovations underscore the transformative potential of combining computational prowess with clinical expertise in battling complex diseases like HCC.

Subject of Research: Microvascular invasion prediction in hepatocellular carcinoma using advanced deep learning models applied to gadoxetic acid-enhanced MRI.

Article Title: Deep multi-instance learning model based on gadoxetic acid-enhanced MRI for predicting microvascular invasion of hepatocellular carcinoma: a multicenter, retrospective study.

Article References: Luo, Y., Zhang, G., Zhong, S. et al. Deep multi-instance learning model based on gadoxetic acid-enhanced MRI for predicting microvascular invasion of hepatocellular carcinoma: a multicenter, retrospective study. BMC Cancer 25, 1626 (2025). https://doi.org/10.1186/s12885-025-14971-7

Image Credits: Scienmag.com

DOI: https://doi.org/10.1186/s12885-025-14971-7