The growing challenge of healthcare-associated infections (HAIs) in the NHS
The scale of the problem: HAIs and NHS capacity
Healthcare-associated infections (HAIs) affect over 850,000 patients each year in NHS hospitals annually, contributing to 50,000 deaths annually. These infections are linked to longer hospital stays (typically an additional 8 days [3]) and higher care costs [1] – the impact of which is not insignificant at a time when services are already over-stretched.
In turn, lengthy stays also increase the risk of additional infections, associated healthcare conditions, such as pressure sores or falls, a reinforcing cycle of clinical decline, as well as antimicrobial resistance (AMR) – one of the biggest health crises worldwide. With growing pressure on NHS bed occupancy which currently stands at 94.8% [4], a remarkable 15-20% of inpatient beds are occupied by HAIs [1], exacerbating the current elective care backlog and estimated to affect around 7 million patients [5]. Evidence shows that more than 50% of these infections could be prevented through better infection control practices [6] – something often deprioritised due to understaffing and overcrowding.
Despite decades of research, current detection systems fail to identify infection risk early enough to prevent deterioration. With increased pressure on services and AMR soaring, the need for intelligent, predictive systems has never been more urgent. Advances in clinical AI offer a compelling solution: early warning tools that not only detect infection sooner but support proactive clinical decisions before deterioration begins.
Accelerating the AMR crisis
AMR is a serious global health issue, shaped by numerous factors including inappropriate use of antimicrobials. The spread of AMR is rapidly accelerating [7], resulting in longer, harder-to-treat illnesses, higher death rates, and significantly increased treatment costs. By 2050, AMR is predicted to cause more deaths than cancer and diabetes combined [8]. There is a strong association between HAIs and AMR, as infections acquired in healthcare settings are more likely to be caused by resistant pathogens. The fight against AMR therefore requires a coordinated, cross-sectoral approach. Whilst public education is essential to promote awareness about the appropriate use of antimicrobials, healthcare systems, including the NHS, must be equipped with advanced diagnostic tools to effectively fight AMR.
Missed opportunities: System shortfalls
The primary tool used for detecting deterioration in patients, including infection-related deterioration, in UK hospitals is the National Early Warning Score v02 (NEWS2) [9]. This rule-based system evaluates six key physiological parameters, including respiratory rate, temperature, and oxygen saturation, to detect acute deterioration. However, NEWS2’s reliance on single-point measurements means it cannot track the trends or changes over time that are crucial for early detection of infections. This limitation can lead to delays in identifying infections, especially in patients who may exhibit atypical vital signs and struggle to communicate their symptoms [10]. While NEWS2 can identify general clinical deterioration, it lacks the precision needed to detect infections specifically. As a result, infections are often detected only after the patient’s condition has already deteriorated, at which point opportunities for early intervention are missed with potentially fatal consequences.
Catching it early: The clinical and operational benefits of early HAI detection
Rapid and accurate identification of infections allows clinicians to prescribe the most appropriate treatment sooner, improving patient outcomes and preserving the effectiveness of existing antimicrobials and reducing the rise of antimicrobial resistant pathogens. Early intervention in infection cases has been shown to reduce high-acuity transfers, decrease sepsis rates [11], and lower the incidence of other healthcare-acquired conditions [5]. From a hospital operations perspective, it will help reduce infection-related bed occupancy [6], thereby improving patient flow and contributing towards reducing the elective care backlog. AI-driven patient risk stratification can help overstretched teams prioritise high-risk cases, reducing clinical teams’ cognitive burden and improving overall efficiency.
Harnessing the latest advancements in AI to improve healthcare-associated infection management
While using healthcare data to develop intelligent tools presents numerous challenges, recent advances in clinical AI have made it possible to develop increasingly sophisticated Early Warning Systems (EWS) and risk stratification tools capable of identifying infections, sepsis, and other adverse conditions. However, there are numerous traditional challenges with healthcare data that these tools must overcome to benefit clinicians and patients:
1. Beyond the Numbers: The power of multimodal data in predicting infection Most early models relied on a single data stream, such as vital signs, missing the richer clinical context available from combining physiological measurements, free-text notes, images and lab results. In the context of infection risk prediction, using multimodal data can offer a more comprehensive view than relying solely on vital signs or lab results, which are often more effective at detecting active infections than predicting them early. For example, researchers at Columbia University Irving Medical Center [12] found that nursing behaviour, specifically, the tendency of experienced nurses to write longer, more detailed notes when concerned about a patient, was a strong early indicator of infection risk.
2. The data dilemma: Unmasking the challenges of real-world data
Patient data is often unlabelled, inconsistent and noisy. Large, accurately labelled datasets are hard to find and expensive to generate. When it comes to infections, this challenge is particularly pronounced, as infections are rarely recorded in a structured or consistent way. Instead, researchers must rely on indirect indicators, such as antibiotic prescriptions or clinical documentation, to infer infection events. Identifying proxies that are both accurate and unbiased is difficult, and poor choices can lead to misleading labels and reduced model performance.
3. Location matters: How data variability impacts AI’s reach
Healthcare data can vary significantly in structure, the way or how often it is recorded across different care settings. Further, demographics, comorbidities and disease prevalence vary by region and care settings, challenging the transferability of models trained in one environment to another. Risk factors for infection can vary widely between younger, urban populations and older, rural ones. Socioeconomic context also plays a key role: models developed using data from wealthier communities may fail to generalise in more deprived areas, where social determinants of health significantly influence infection risk. These discrepancies make it difficult to standardise datasets for training and can degrade model performance upon deployment.
There have been a number of recent breakthroughs in clinical AI that address these obstacles:
- Deeper clinical insights with multimodal AI
Recent progress in pre-training techniques, such as self-supervised learning, and in feature fusion methods is helping address the challenges of working with multimodal data. These approaches allow models to process and combine information from different data streams in a way that preserves the most relevant and complementary signals from each. Rather than treating each data type in isolation, these methods help the model learn a shared representation that captures the full clinical picture, improving its ability to detect subtle patterns.
- Self-supervised Learning (SSL) and healthcare data:
This allows models to learn meaningful representations from vast unlabelled datasets, mirroring approaches used in large language models, so they can infer relationships (for example, between antibiotic orders and infection indicators) without requiring manual annotation.
- Domain adaptation and continuous learning
This provides the flexibility to acknowledge the changing nature of data in different care settings, different patient populations, and different recording practices etc. Domain adaptation allows models to adjust to these differences when retraining on a new dataset by learning a shared feature space. Simultaneously, continuous learning prevents “catastrophic forgetting” (forgetting everything learned from previous datasets) by replaying samples of older datasets during retraining.
Overcoming these data-centric hurdles demands a new generation of AI advancements, which recent breakthroughs in clinical AI are now poised to deliver. To ensure that AI tools are actively used by clinical teams, four critical requirements must be met:
- Seamless integration into hospital systems: To minimise additional cognitive burden on already overstretched staff, AI tools must plug directly into existing electronic patient record (EPR) systems. Embedding alerts, risk scores and decision-support prompts within familiar workflows prevents tool fatigue and supports timely, context-aware interventions.
- Codesign with clinical teams: AI solutions must be both useful and usable, which means building them hand-in-hand with the end users. This collaboration should not limit itself to user interface design, but it should also touch upon key product considerations such as setting up alert thresholds. For instance, in infection prediction, the risk threshold for flagging a patient as “high risk” will vary by patient group and care setting. In highly vulnerable populations, where missing a single infection could have catastrophic consequences, clinicians may accept more false positives, so no true infections are overlooked. Ultimately, the real value of AI lies not in the algorithm’s score itself, but in the clinical decisions it enables.
- Trust is driven by explainability: Clinicians must understand why and how a model generates each prediction. Explainable AI not only fosters confidence in the tool but also allows teams to identify and correct model shortcomings. For example, if a particular feature carries disproportionate weight in the model compared to its real-world significance, clinicians can flag this and guide its recalibration.
- Regulatory approval as Software as a Medical Device: Any AI tool employed in patient care must satisfy the same safety and efficacy standards as traditional medical devices. Demonstrating compliance with regulatory requirements reassures both clinicians and patients that the technology meets rigorous benchmarks for clinical use.
Reimagining healthcare delivery: AI as a preventative tool
The NHS, like many healthcare systems around the world, is facing unsustainable pressure, with demand consistently outstripping capacity and care often delivered too late. Healthcare-associated infections are a clear example of this reactive model: when caught early, treatment can be more targeted and less invasive, and timely isolation can prevent hospital-wide outbreaks. Yet in many cases, infections are detected too late, resulting in rushed interventions, broad-spectrum antibiotic use (which increases the risk of antimicrobial resistance), and serious complications for patients, including ICU admissions.
Today, the widespread adoption of electronic patient records (EPRs) provides access to rich, multimodal datasets that can power the next generation of early warning systems and risk stratification tools. Recent advances in clinical AI, such as self-supervised pre-training, multimodal fusion, domain adaptation, and continuous learning, are beginning to overcome long-standing technical barriers
However, technology alone is not enough. The most impactful AI solutions will emerge from close partnership with frontline clinicians, ensuring that models reflect real-world workflows and clinical judgement. A rigorous evaluation framework, ideally driven by public and independent bodies such as NICE, will be vital to establish both safety and true clinical benefit, avoiding pitfalls like “alert fatigue” or the Early Warning System paradox [14]. By combining data-driven innovation with practical, clinician-led design and robust assessment, we can begin to move from a reactive to a preventative healthcare model, and, in doing so, improve outcomes for patients across the NHS and beyond.
