Device manufacturers that unravel technical and regulatory issues will lead the fight.
By Rob Morgan, PhD, Paulo Pinheiro, PhD, and Andrew Chapman, MBA
Hospital-acquired infections, also known as healthcare-associated infections (HAIs), are a major burden for global healthcare. In the United Kingdom, they represent an annual cost of around £1billion to the National Health Service.1 And in the United States, the Centers for Disease Control and Prevention estimates that HAIs account for 1.7 million infections and 99,000 associated deaths each year.2 HAIs can include a number of conditions, such as hospital-acquired pneumonia, catheter-associated urinary tract infections, and surgical site infections, which may develop in patients who have undergone invasive treatments or procedures involving devices, surgical equipment, or monitoring technologies. The same equipment that has revolutionized healthcare—improving patient outcomes and saving lives—can sometimes result in life-threatening aftereffects. Inadequate cleaning, maintenance, or design of this equipment can have fatal consequences. Sepsis is one of the most prolific and deadly HAIs. Worldwide, 11 million sepsis deaths were reported in 2017, representing 19.7% of all global deaths, according to a report published in Lancet.3 The “global burden of sepsis is larger than previously appreciated, requiring further attention,” the authors concluded. Furthermore, they clearly highlight a link with HAIs: “Many of these cases of sepsis are suspected to be due to nosocomial infections; patients admitted to hospital for noninfectious conditions could be exposed to infection risk either from invasive devices such as central venous or urinary catheters or through inadequate handwashing practices among healthcare workers.” The study’s authors conclude that clinicians and public health policy makers must implement cost-effective measures to improve sepsis outcomes. And one factor critical to achieving that goal is early detection and treatment. However, early detection is a major challenge with sepsis, so medical device and healthcare providers are increasingly looking at how clinical decision support (CDS) software can incorporate artificial intelligence (AI) and machine learning to augment and accelerate the diagnostic process for this disease.
Challenge of Sepsis Diagnosis
Diagnostic challenges impede the fight against sepsis. In its early stages, the disease is easy to treat but hard to diagnose; in later stages diagnosis is easy but treatment is harder. Key early symptoms of sepsis—low blood pressure, elevated white blood cell count, an increased heart rate, and shallow breathing—are standard for any infection. These signs alone in a patient would not necessarily prompt healthcare professionals to request scans or screen for acid or bacteria in blood or other bodily fluids. But sepsis progresses very quickly, and every hour that treatment is delayed reduces the average survival rate by 8%. By the time sepsis is diagnosed, it can be too late to prevent tissue damage, organ failure, or death. Clearly, it’s not practical or desirable to screen every patient who presents these early symptoms for sepsis. However, better understanding of factors that might increase the likelihood that a patient could be developing sepsis—for example, after invasive procedures—would be hugely beneficial. A targeted approach that optimizes the timing and workflow of diagnosis and treatment would mean that patients who do have sepsis are identified and receive antibiotic treatment in those critical early hours when a positive outcome is more likely. A fine balance needs to be achieved: Early prediction of sepsis can be lifesaving, but screening for sepsis in every patient with signs of infection risks wasting valuable hospital resources.
AI Optimizes the Diagnostic Pathway
Machine learning and AI offer new ways to overcome the challenge of diagnosing sepsis early with intelligence-led predictions of a patient’s likelihood of developing the disease. CDS software with this capability could potentially drive significant improvements in patient outcomes. Modern healthcare generates huge amounts of data—including each patient’s vital signs, lab test results, progress notes, and medications—but often the information is held in disparate and remote systems. Applying machine learning and AI to this big data can yield important insights and predictive capabilities that have previously been out of reach. These predictive capabilities can be leveraged via CDS software to drive benefits in three core areas: • Diagnosis. Predictive capabilities enable earlier diagnosis and intervention when a patient has sepsis. • Prognosis. Predicting readmission due to sepsis can ensure patients most at risk are identified and closely monitored. • Treatment. Data surrounding treatments and outcomes can be harnessed to devise optimal treatment strategies for sepsis in intensive care. The wider medical community has expressed interest in the use of data for improved and predictive sepsis diagnosis, and pockets of activity in various fields of medicine are already making some headway. For instance, the PhysioNet Computing in Cardiology Challenge 2019 focused on early prediction of sepsis using clinical data. Participants were challenged to accurately predict sepsis using physiological data 6 hours ahead of a clinical prediction. A report discussing the outcomes of the challenge concludes that “Diverse computational approaches predict the onset of sepsis several hours before clinical recognition, but generalizability to different hospital systems remains a challenge.”4
Advancements in Healthcare Technologies
Several large healthcare providers and electronic health record (EHR) software providers are integrating AI and machine learning with their efforts to reduce sepsis-related mortality. A collaborative project between Geisinger Health System and IBM analyzed deidentified EHR data for more than 10,500 sepsis patients in the United States. Their efforts resulted in the creation of an AI model to predict sepsis mortality that will aid the development of personalized clinical care plans for at-risk patients. Additional developments in this space range from sepsis prediction tools to real-time evaluation of patient condition: • SPOT (sepsis prediction and optimization of therapy) is an electronic information and alert system developed by HCA Healthcare. Embedded in a patient’s EHR, SPOT analyzes real-time data from bedside monitoring equipment and medical lab test results. HCA claims SPOT can identify sepsis approximately 18 hours earlier than the best clinicians, and the system alerts physicians and caregivers accordingly. • Sepsis Watch, developed by Duke Institute for Health Innovation, is an AI-enabled system that pulls information from a patient’s EHR every 5 minutes to evaluate condition and offer real-time analysis that human doctors cannot provide. • Jvion, a healthcare AI company, plots individual patients on a sepsis risk trajectory. It then determines whether the trajectory can be changed and provides patient-specific recommendations. EHR specialists EPIC and Cerner have also developed sepsis predictive tools and alerts that are gaining attention in the United States. These developments are promising, and progress is certainly being made. However, assessing the true value of any AI-informed decision is challenging. When the algorithm and clinicians disagree, it’s difficult to reliably estimate what would have happened to the patient in an alternate reality. Furthermore, achieving meaningful improvements in nosocomial sepsis outcomes at scale will require a more widespread, joined-up approach than we have seen to date.
Barriers to AI-Led Sepsis Diagnosis
Currently, many obstacles—both technical and regulatory—are preventing the large-scale use of AI and machine learning in CDS software to support the diagnosis of sepsis. Technical issues range from insufficient data standardization and integration to poor interdevice communication in healthcare settings. Standardization of data is a major challenge preventing scaled use of AI for predictive sepsis diagnosis. Overcoming this obstacle is critical so that data can be converted into a common format that is understood across multiple implementations. Interoperability between different workflow components is also essential to facilitate the storage and retrieval of data from EHRs. A related issue is the need for external validation of AI algorithms, a process that demands access to and interoperability between different data sets by an independent party. Randomized controlled trials comparing “clinicians alone” to “clinicians assisted by the algorithm” may be necessary for regulatory approval. The data set shift phenomenon also needs to be considered. An AI algorithm is typically “trained” within a stationary environment. Introducing AI to sepsis management is likely to cause changes in practice, which in turn will result in a new distribution, different to that used when training the algorithm. To counter this shift, methods need to be put in place to identify performance deterioration. One of the greatest challenges of AI is reliable generalization. Generalization can be hard due to technical differences between hospitals (including differences in equipment, coding definitions, and EHR systems as well as laboratory equipment and assays). Variations in local clinical and administrative practices and the population itself are another factor. So site-specific training will be required to adapt existing systems for new populations. Methods to detect out-of-distribution inputs and provide a reliable measure of model confidence will be important to avoid clinical decisions based on inaccurate model outputs. Communication between devices provided by different vendors and where AI algorithms are hosted is another important concern. This is particularly true for sepsis, where AI diagnosis tests may be conducted at 5-minute intervals. It’s likely that many manufacturers will aim to make their product the linchpin of a connected AI ecosystem, drawing in data from other devices. So common standards for good device integration will be essential. Regulatory matters need to be considered at an early stage of device development. In 2019, the FDA released new draft guidance on CDS software.5 The guidance includes a category dedicated to “Device CDS,” which uses the same risk classification framework as “Software as a Medical Device.” While the final details are yet to be announced, the FDA clearly indicates an intention to focus its regulatory oversight on Device CDS functions for healthcare practitioners (HCPs) that “‘inform clinical management’ for ‘serious or critical situations or conditions’ and that, in addition, are not intended for the HCP to be able to independently evaluate the basis for the software’s recommendations.” This statement suggests that the use of AI and machine learning in devices geared toward the diagnosis of sepsis will be subject to intense scrutiny. In addition, Article 22 of the European Union’s General Data Protection Regulation (GDPR), which gives people the right to receive an explanation for algorithmic decisions, must be considered. This requirement potentially limits the deployment of devices if people operating them cannot understand or interpret how the AI algorithm reached a certain decision or prediction. These technical challenges and regulatory matters are not insurmountable. However, addressing them effectively will require focused attention and collaborative effort from a broad range of specialists, including experts in AI, machine learning, software development and data, as well as healthcare. Pooling all of this expertise and insight could enable CDS software to achieve the required standards for improved nosocomial sepsis diagnosis and treatment. Future of Sepsis Management AI and machine learning present an exciting opportunity to facilitate the early detection and treatment of hospital-acquired sepsis. A broad spectrum of equipment manufacturers has opportunities to optimize or future-proof products so they can play an active role in sepsis management. Those on the hardware side of the healthcare ecosystem might obtain more frequent and automated monitoring of vital signs, which can be uploaded to EHRs and accessed by AI tools. Many in vitro diagnostic companies are already investing in better and quicker infection testing, so embracing predictive sepsis diagnosis could be a natural next step. Predictive diagnosis can be hugely beneficial, even if it doesn’t provide a firm diagnosis. For instance, a predication might prompt a caregiver to run an initial blood or urine test to establish whether an infection is present before confirming whether it is sepsis. Since the survival rate is 80% when treatment is administered within the first hour, starting a patient on a broad-spectrum antibiotic could make a significant difference to the outcome. More targeted antibiotic therapy could be administered once confirmatory test results identifying the infectious organism are obtained.
In time, patients with sepsis will be optimally managed by a combination of AI algorithms and human clinicians working hand in hand. The same will be true of other infections and health issues related to hospital treatment, from postoperative cardiovascular events to opioid addiction. Medical device companies that act fast to address the technical and regulatory factors limiting AI and machine learning will be at the forefront of this new reality.
Rob Morgan, PhD, is vice president of Medical at Sagentia. With 20 years’ experience in medical technology development, he works with clients across the medical devices and pharmaceutical industries. Areas of focus include the development of surgical systems, in vitro diagnostics, digital health solutions, and drug delivery devices. Paulo Pinheiro, PhD, is head of Electronics, Software and Systems at Sagentia. Specialist areas include minimally invasive robotics, cost reduction of blood diagnostics, handheld surgical devices, medical grade bioinformatic pipelines for DNA testing, mobile medical apps, and many more.
Andrew Chapman, MBA, is vice president of Medical at Sagentia. During his career, he has created breakthrough physical and digital products for some of the world’s largest companies. He works with pharmaceutical and med tech clients to develop disruptive new medical technologies for deployment in various healthcare settings. References 1. Funding to fight hospital acquired infections. Lifescience Industry. https://www.lifescienceindustrynews.com/money/funding-to-fight-hospital-acquired-infections/. Accessed August 10, 2020. 2. Healthcare-acquired infections (HAIs). Patient CareLink. https://patientcarelink.org/improving-patient-care/ healthcare-acquired-infections-hais/. Accessed August 10, 2020. 3. Rudd KE, Johnson SC, Agesa KM, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study. Lancet. 2020;395(10219):200-211. doi: 10.1016/S0140-6736(19)32989-7. 4. Reyna M, Josef C, Jeter R, et al. Early prediction of sepsis from clinical data: the PhysioNet/Computing in Cardiology Challenge 2019. Crit Care Med. 2020;48(2):210-217. doi: 10.1097/CCM.0000000000004145. Available at https://www.physionet.org/content/challenge-2019/1.0.0/physionet_challenge_2019_ccm_manuscript.pdf. 5. Clinical decision support software. Food and Drug Administration. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software. September 2019. Accessed August 11, 2020.