The global public health system is confronting a slow-burning but catastrophic crisis as the very medicines designed to save us are losing their power against an invisible enemy. This silent pandemic of antimicrobial resistance (AMR) threatens to unwind a century of medical progress, turning routine infections into life-threatening events. In a landmark initiative to turn the tide, Singapore General Hospital (SGH) has embarked on a pioneering collaboration with DXC Technology, harnessing the formidable capabilities of artificial intelligence to forge a new line of defense. This project moves beyond theoretical applications, creating a tangible, sophisticated AI system engineered to revolutionize the hospital’s Antimicrobial Stewardship Program (ASP). Traditionally a manual and labor-intensive process, antimicrobial stewardship is being reimagined as an automated, data-driven powerhouse. By developing an AI model capable of auditing antibiotic prescriptions with unparalleled speed and precision, the collaboration aims to arm medical teams with the actionable insights needed to optimize treatments, curb the emergence of drug-resistant “superbugs,” and ultimately secure better patient outcomes through reduced mortality, shorter hospital stays, and lower readmission rates.
The Growing Crisis and the Limits of Traditional Defenses
The Superbug Threat
The pervasive overuse and misuse of antibiotics in clinical settings represent a critical catalyst for the rise of superbugs, with authoritative estimates suggesting that a staggering 20 to 50 percent of all antibiotic prescriptions within acute care facilities are either unnecessary or inappropriate. This widespread practice provides the ideal evolutionary pressure for microorganisms to develop resistance, rendering our most reliable treatments ineffective. In Singapore, the challenge is particularly pronounced; heavy reliance on antibiotics has fueled significant resistance in Gram-negative bacteria, and an alarming 30 percent of hospital-acquired infections now exhibit resistance to widely used broad-spectrum drugs. This escalating resistance transforms hospitals from places of healing into potential breeding grounds for formidable pathogens, where patients seeking treatment for one condition can acquire another that is far more difficult to manage. The consequences extend beyond individual patient outcomes, placing immense strain on healthcare systems and driving up medical costs substantially.
Compounding this dire situation is the stark reality of a near-empty pipeline for new antibiotics. The economic and scientific hurdles of drug development have led pharmaceutical innovation in this area to a virtual standstill. As existing antibiotics progressively lose their efficacy, the absence of novel replacements creates a perilous void in our therapeutic arsenal, threatening to plunge medicine back into a pre-antibiotic era where common injuries and minor surgeries could lead to fatal infections. This reality elevates the responsible management of our current antibiotic supply from a recommended best practice to an absolute imperative for preserving the foundations of modern healthcare. Without effective antimicrobial agents, medical advancements such as organ transplantation, chemotherapy, and complex surgical procedures would become fraught with unacceptable risk, fundamentally altering the landscape of patient care and public health safety. The fight against AMR is therefore not just about controlling infections; it is about safeguarding the entirety of our medical infrastructure.
The Challenge of Manual Stewardship
Antimicrobial Stewardship Programs (ASPs) have emerged as the primary strategic defense against the escalating threat of AMR. These programs, which are well-established and rigorously implemented across Singapore’s public hospitals, have demonstrated significant success in their mission. Through meticulous oversight and intervention, ASPs have been instrumental in curbing the misuse of antibiotics, leading to a cascade of proven benefits for both patients and the healthcare system. Clinical data consistently shows that effective stewardship directly contributes to shorter hospitalizations, reduced patient mortality rates, and substantial cost savings. By ensuring that antibiotics are prescribed only when necessary, in the correct dosage, and for the appropriate duration, these programs help slow the development of resistance while optimizing the care provided to each individual, reinforcing their role as a cornerstone of modern infectious disease management.
Despite their proven value, traditional ASPs are constrained by a fundamental limitation: the overwhelming scale of the problem far exceeds the capacity of manual human review. The experience of SGH in 2018 serves as a powerful case study, where a dedicated team conducted approximately 12,000 manual audits of just seven broad-spectrum intravenous antibiotics. This exhaustive effort revealed that a concerning 20 to 30 percent of these prescriptions were inappropriate. The immense human capital required to perform this level of scrutiny is staggering, and the prospect of expanding such a detailed review to encompass the full spectrum of antibiotics—including both broad and narrow-spectrum agents, administered intravenously or orally—is simply not feasible. This scalability issue creates a critical vulnerability in our defenses, as the vast majority of prescriptions go unaudited, leaving significant room for inappropriate use to continue unchecked and fuel the engine of antimicrobial resistance.
A New Strategy: AI-Powered Pre-emptive Care
Shifting from Reactive to Proactive
The inherent limitations of manual auditing have created an urgent need for a fundamental paradigm shift in antimicrobial stewardship, moving from a largely reactive model to a pre-emptive and proactive strategy. In the current standard of practice, the review of an antibiotic prescription often occurs several days after a patient has already begun the course of treatment. This delay means that any intervention to correct an inappropriate prescription is inherently retrospective, and the patient may have already been exposed to an unnecessary or suboptimal drug, contributing to both individual risk and the broader pressure for resistance. The innovative AI-driven model developed by SGH and DXC Technology aims to dismantle this outdated approach by integrating stewardship directly into the initial decision-making process, empowering physicians with data-driven guidance at the point of prescription to ensure the most appropriate treatment is selected from the outset.
To prove the efficacy of this new pre-emptive model, pneumonia was strategically selected as the ideal starting point for its implementation. This decision was based on several key factors that make pneumonia a particularly compelling use case. It is a highly prevalent condition, accounting for approximately 20 percent of all infections treated at SGH, and it stands as the leading diagnosis for which antibiotics are prescribed. Furthermore, the complexity of pneumonia diagnosis underscores the need for a more precise, data-driven methodology. The condition can be viral, in which case antibiotics are completely ineffective, or it can be bacterial, requiring a carefully chosen antibiotic. Even when bacterial, the infection may be treatable with oral drugs rather than more potent intravenous ones. This diagnostic ambiguity often leads to the precautionary use of broad-spectrum antibiotics, making it a perfect testing ground to demonstrate how a sophisticated, patient-centered AI can guide clinicians toward more precise and effective treatment decisions.
The AI’s Clinical Reasoning Framework
At the core of this forward-thinking strategy is a comprehensive AI solution built upon a detailed clinical reasoning framework designed to systematically emulate and augment the complex diagnostic process of an experienced physician. This framework is meticulously structured to guide clinicians toward an accurate diagnosis and, subsequently, the optimal treatment plan. It operates by systematically posing and answering five categories of essential questions, each drawing upon a vast and diverse array of patient data to construct a holistic clinical picture. The foundational inquiry, “Is there an infection?”, prompts the AI to analyze a multitude of biological markers to confirm the presence of an infectious process and, critically, to differentiate its nature—for example, distinguishing a bacterial infection that requires antibiotics from a viral one that does not. This initial step involves processing data points such as the patient’s hemodynamic status, fever, results from blood tests like white blood cell and neutrophil counts, sputum cultures, and viral panel screenings.
Following the confirmation of an infection, the framework proceeds through a logical sequence of inquiries to refine the diagnosis and treatment plan. The second question, “Where is the infection?”, focuses on localizing the source by interpreting the patient’s presenting complaints alongside biomarkers from laboratory tests and findings from X-rays and other imaging studies. Next, the question “When did the infection occur?” helps determine whether it is community-acquired (presenting within 48 hours of hospital admission) or hospital-acquired (developing after 48 hours), a distinction with significant implications for the likely pathogens and appropriate treatment. The framework then turns to the patient with “Who is the host?”, considering their individual profile, including medical comorbidities, recent hospitalizations, and known allergies. Finally, with the answers from the preceding steps narrowing the therapeutic options, the AI addresses the ultimate question: “What antibiotic should be selected?” This final step synthesizes all available data to refine the choice, considering factors such as the presence of systemic inflammatory response syndrome (SIRS), whether the patient is in intensive care, their recent antibiotic history, and the results of any culture sensitivity reports.
Overcoming Data Hurdles
A primary challenge in operationalizing this sophisticated clinical reasoning framework was the immense technical complexity of managing the disparate and fragmented data ecosystem within the hospital. To function effectively, the AI model must draw information from a multitude of siloed sources, including the institution’s proprietary electronic medical records (EMR) system, linked data repositories containing radiology images, comprehensive laboratory results, histology reports, and even anesthetic charts from surgical procedures. Integrating these varied data streams into a single, cohesive analytical pipeline presented significant technical hurdles related to data quality, engineering, freshness, and completeness. The raw data exists in numerous formats, with inconsistencies and gaps that could easily mislead an algorithm if not properly addressed, making robust data management the bedrock of the entire project.
To surmount these obstacles, the project team devised and implemented a series of sophisticated strategies aimed at transforming raw medical information into analytically useful components. A crucial element was advanced data engineering, which involved developing processes to distinguish the lifelong relevance of a chronic illness from the short-term validity of a single lab test result, ensuring that the AI could correctly weigh the importance of different data points over time. Furthermore, the team employed data-completeness techniques to intelligently handle instances of missing information. Rather than discarding an incomplete patient record, the system was designed to compute proxy values based on other available medical factors or to use statistical methods like averaging to fill in gaps, thereby preserving the integrity of the dataset. These meticulous efforts ensured that the AI model was trained on high-quality, comprehensive data, enabling it to generate reliable and clinically relevant insights.
The AI in Action: Promising Early Results and Future Vision
Performance and Efficiency Gains
The intensive collaboration between clinical experts and data scientists has yielded two distinct AI models, both of which have demonstrated promising initial results that validate the project’s ambitious goals. The first, an Operational AI Model, was specifically designed to act as an intelligent filter, identifying patient cases that warrant further review and thereby prioritizing the ASP team’s limited time and resources. In a rigorous test involving 2,012 pneumonia treatments that the model had never seen before, it showcased remarkable efficiency. While the standard manual process would require reviewing all cases to find the 4% that need intervention, the AI model flagged just 624 cases (approximately 31%) for immediate attention. Critically, within this much smaller, prioritized subset, it successfully captured 72 of the 82 total intervention cases, achieving an impressive 87% capture rate. This effectively reduces the team’s review workload by a factor of three while simultaneously increasing the probability of identifying a case needing intervention from a baseline of 4% to 11.5%—a threefold increase in focus and efficiency.
The second model developed is a Recommender AI Model, which takes the process a step further by aiming to suggest the appropriate antibiotic regimen for cases that have been deemed inappropriate. This model faced a greater degree of complexity due to the vast number of possible antibiotic combinations and the nuanced clinical factors that guide such decisions. Despite these challenges, it was able to recommend the correct antibiotic regimen in 50% of the cases it analyzed. Beyond sheer accuracy, the transformative speed of the AI is a game-changing factor. A comprehensive manual case review typically takes a skilled clinician around 20 minutes to complete. In stark contrast, the AI can assess the same complex web of data in less than a second. This incredible velocity enables real-time analysis, allowing the ASP team to review a much larger volume of prescriptions and expand its vital oversight to a wider set of antibiotics, fundamentally altering the scale and immediacy of antimicrobial stewardship.
The Path Forward
The successful implementation of these AI models was not an endpoint but rather a foundational step in a much broader vision for the future of medicine. This project has set the stage for a multi-phase progression where the integration of advanced data analytics and artificial intelligence into clinical workflows becomes standard practice, augmenting human expertise rather than replacing it. The immediate next phase will focus on gathering deeper insights across the full spectrum of antibiotics used to treat pneumonia, further refining the models’ accuracy and expanding their knowledge base. Following that, the third phase aims to scale the model’s capabilities to cover all infectious diseases treated within the hospital, creating a comprehensive, AI-powered stewardship system. The ultimate objective is to have this automated system fully integrated with the hospital’s core clinical systems, providing real-time decision support to clinicians directly at the point of care.
This final integration would enable the prevention of inappropriate antibiotic use from the moment of prescription and facilitate the development of highly personalized treatment regimens tailored to each patient’s unique biomarkers, symptoms, comorbidities, and medical history. Achieving this ambitious goal has underscored the need for access to even broader and more interconnected datasets, potentially including community health data to provide a more complete picture of a patient’s risk factors. The implications of this work extended far beyond a single institution. As a recognized center of excellence, Singapore’s leadership in tackling AMR through this innovative fusion of AI and clinical stewardship has established a powerful precedent. This initiative has served as a blueprint for other healthcare systems worldwide, demonstrating a viable, scalable, and highly effective strategy that has enhanced the situational awareness of prescription practices and advanced the critical cause of global antimicrobial stewardship.
