AI helping detect sepsis in the near future

A surge of information

Modern healthcare is replete with digital footprints. Today, our hospitals and clinics are more than just buildings with doctors and patients; they’re hubs of vast amounts of electronic data, and modern medicine is looking at ways to store and use this data in real time. From vital sign monitors to laboratory test results, and from progress notes to intricate billing data, healthcare systems are generating and storing a monumental amount of information personal to each patient journey. While this avalanche of data promises unmatched insights into patient health and medical patterns, it also presents an overwhelming challenge for physicians. The task of assimilating this data, understanding its implications, and then making informed decisions is becoming increasingly daunting. Amidst this backdrop, the emergence of Artificial Intelligence (AI) offers a glimmer of hope.

Though AI often makes headlines in popular culture, its practical applications in medicine are both scary, groundbreaking and transformative. AI aims to imbue machines with capabilities that have historically been the domain of human intellect, such as reasoning and decision-making. We have already seen the impact AI can have in assisting clinical decision making through radiography, cellular science and even in the detection of cancer. 

A refresher of how AI learns in medicine

At the heart of this AI revolution is a sub-discipline known as machine learning, which has exhibited significant promise in revolutionising medical diagnoses, prognoses, and treatment recommendations. Broadly, there are two main machine learning methodologies finding traction in healthcare:

Supervised Learning: Here, models are meticulously trained using known data, enabling them to make informed predictions when faced with new, sometimes ambiguous, data. 

Reinforcement Learning: This approach is centred around experimentation. The model refines its decision-making prowess through a series of actions, assessing the outcomes of each, and adapting accordingly. It’s akin to a chess player strategizing several moves in advance, considering both the immediate consequences and the long-term repercussions of their choices.

AI shines brightest when confronted with multifaceted scenarios, characterised by a myriad of interplaying factors. In these cases, traditional analytical methods (and the human subjective nature) might falter, but AI models may unearth surprising interactions and correlations. The complex domain of sepsis serves as a pertinent example.

The complication of sepsis

Sepsis stands as one of the most daunting challenges in contemporary healthcare. Its severe nature demands timely detection and intervention. As per the sepsis-3-criteria, sepsis manifests as a significant rise in the Sequential Organ Failure Assessment (SOFA) score, which points to potential life-threatening organ dysfunctions triggered by infections.

In its nascent stages, sepsis is relatively manageable with treatments like broad-spectrum antibiotics. Yet, its initial symptoms are often elusive, overshadowed by complicating and often serious co-morbidities, masking its presence. As the condition advances, its detection becomes clearer, but the treatment landscape gets murkier, often with dire consequences.

Complicating the clinical landscape further is sepsis’s intrinsic heterogeneity. It doesn’t follow a one-size-fits-all pattern. Patients might develop sepsis from an array of pathophysiological mechanisms, each presenting with unique clinical phenotypes. Alarmingly, approximately 20% of patients who report to emergency departments with potential sepsis don’t exhibit immediate signs of organ dysfunction. However, this can change dramatically within just 48 hours. While we have some tools like the quick Sequential Organ Failure Assessment (qSOFA) at our disposal, their sensitivity is often not up to the mark, leading to missed or delayed diagnoses by a range of practitioners along the chain.

Sepsis detection assisted with AI

A state-of-the-art AI system has been devised at Johns Hopkins University, showing remarkable efficacy in the early detection of sepsis symptoms. This advancement comes after an extensive study, indicating a 20% reduction in sepsis-related mortality, given that the system can recognize symptoms hours prior to conventional methodologies. The AI model delves into medical records and clinical notes, pinpointing patients who may be on the verge of severe sepsis-related complications. This ability to piece together an entire patient picture holistically is something that AI thankfully does with ease. Suchi Saria, the lead author of the research published in Nature Medicine and Nature Digital Medicine, hailed it as a groundbreaking bedside AI application, one which has been used by thousands of healthcare providers. 

The magnitude of this achievement can’t be overstated, considering sepsis, an infection-induced chain reaction causing potential organ failure, afflicts approximately 1.7 million US adults annually, with a staggering quarter-million succumbing to it. In Australia, the statistics are equally as alarming. An estimated 18,000 Australian adults are treated in intensive care units for sepsis annually, of which a staggering 5,000 on average will die due to these complications.

The “Targeted Real-Time Early Warning System” (TRTEWS), as termed by the Johns Hopkins team, amalgamates a patient’s historical medical data with their current symptoms and lab results. This provides clinicians with real-time insights, suggesting optimal treatment protocols such as antibiotic initiation. A salient feature is the system’s ability to monitor patients continuously, from admission to discharge. 

During its evaluation phase, the AI was used by over 4,000 clinicians across five hospitals to treat 590,000 patients. With an accuracy rate of nearly 40% in 82% of sepsis instances, it vastly outperforms older electronic detection tools, which hovered around a mere 2% to 5% accuracy, reports the team. In the gravest sepsis cases, where every hour’s delay can be fatal, the AI system astonishingly preempted traditional detection methods by an average of six hours. Albert Wu, co-author and director at Johns Hopkins Center for Health Services and Outcomes Research, lauded it as a genuine breakthrough, especially because the AI tool explains its reasoning behind specific medical recommendations, thereby instilling trust in the healthcare providers.

Whilst this digital integration certainly doesn’t remove the need of the physician and team in the diagnosis of sepsis, it does support the decision making process and identify vulnerable patients to this condition. Whilst our hospital system may not be adequately equipped for this technology in Australia just yet, it is promising to see the use of AI angled as an adjunct to physician capabilities. 

For more information on sepsis in Australia and the need to improve community awareness about the condition, see the amazing work of the Australian Sepsis Network here: https://www.australiansepsisnetwork.net.au/ 

https://hub.jhu.edu/2022/07/21/artificial-intelligence-sepsis-detection/