Augmented intelligence in precision medicine: Transforming clinical decision-making with AI/ML and/or quantitative systems pharmacology models

Abstract

Effective disease management is becoming increasingly important given the chronic disease burden, which is projected to reach $47 trillion worldwide by 2030 (see Appendix S1). Lifestyle and drug adherence significantly impact chronic disease management, making the doctor-patient relationship a key driver of clinical outcomes. By effectively communicating with the patients and empowering them to focus on self-care, physicians can enable lifestyle changes and improve drug adherence, leading to better outcomes. Digital health is transforming the doctor-patient relationship,¹ with patients becoming more proactive in their care, seeking to understand different treatment options, and participating in decision-making. This presents both a challenge and an opportunity—a challenge for the physicians to adapt to evolving patient expectations and an opportunity for physicians empowered with digital tools¹ to not only treat the disease effectively but also impact the overall patient journey. Artificial Intelligence, including machine learning algorithms (AI/ML), spans a broad set of digital tools with great potential to aid physicians in this transformation and revolutionize the care paradigm.

Physicians are increasingly employing AI/ML to improve diagnostics, determine disease prognoses, reduce workload, and support clinical decision-making.² Clinical decision-making involves diagnosing based on patient history, physical examination, and diagnostic tests, followed by determining the optimal treatment plan. Physicians integrate information and evidence from multiple sources to make certain decisions. Despite their expertise, physicians may err due to biases or blind spots, potentially leading to misdiagnosis, or poor treatment choices and suboptimal patient outcomes. AI/ML can help physicians cover their blind spots, eliminate potential biases, and enable data-driven decisions. AI/ML models could also serve as tools to engage patients in discussions about their disease state and update treatment plans to meet their health goals.

AI/ML-based personalization of treatment plans can be achieved by clinical decision support systems (CDSS³) powered by models predicting treatment outcomes and risk of complications, and by digital twins⁴ mirroring actual patients. A digital twin emulates the behavior of a physical system, here a patient, using real-time data to update itself through its lifecycle. The primary differentiator for digital twin platforms from other predictive models is that the digital twins have a memory of the patient's history, evolve with them, and guide the patient toward their goal. In the following sections, we will explore how predictive models and digital twin platforms personalize treatment plans.

Predictive modeling has immense potential to transform treatment decision-making across various disease domains. AI and ML algorithms analyze patient data to provide insights into disease progression, treatment response, and potential outcomes, enabling personalized surgeries and tailored treatment plans. For example, the Virtual Epileptic Patient (VEP) framework is being evaluated in an ongoing clinical trial (EPINOV: NCT03643016)⁵ to estimate its impact on surgical prognosis in epilepsy patients. In pharmacotherapy, predictive models that forecast treatment outcomes in Major Depressive Disorder (MDD) patients can help clinicians identify suitable interventions for individual patients, potentially reducing the duration of depressive episodes and preventing severe complications such as suicide.⁶

Similarly, patients with Ulcerative Colitis face a wide range of treatment choices, yet there is a lack of definitive guidance for selecting the most appropriate option. Predictive models can aid physicians in determining the optimal treatment paths by assessing factors such as disease severity, patient preferences, and potential side effects. This personalized approach can improve treatment efficacy, enhance patient’s quality of life, and reduce healthcare costs.⁷ In oncology, predictive models have demonstrated significant promise. For example, AI-powered tools can analyze patient data to predict the likelihood of hepatocellular carcinoma recurrence following surgical resection, enabling tailored surveillance and preventive measures.⁸ Additional examples of AI/ML-based CDSS can be found in the review article by Sutton et al.³

In the examples above, clinicians input historical data and clinical variables into a model to receive model-based insights for treatment decisions. However, these systems lack the memory of individual patients. In contrast, a digital twin-powered system has memory and feedback, owing to the bi-directional transfer⁴ of information between the actual patient and their digital twin. In a CDSS powered by digital twins,⁹ each patient is assigned a unique digital twin during their initial clinic visit (Figure 1). The digital twin predicts disease trajectory and provides insights into treatment options to meet patient's goals. As new data is collected, it is entered into the digital twin system to compare the digital twin predictions to actual outcomes. If prediction errors exceed an acceptable threshold, the digital twin updates with the new data, offering new insights that may inform the treatment strategy. Through this bi-directional informational transfer, represented by red and blue arrows in Figure 1, a digital twin evolves with time and guides the individual patient toward optimally achieving their health goal, thus assisting the physician in the process. Figure 1 also illustrates a scenario where a patient either does not meet their treatment goal or takes longer as their physician lacks additional insights to provide an optimal treatment plan.

A fundamental component of digital twin creation is a model that accurately represents the system's essential dynamics, whether of a single organ or a complex disease involving multiple organs. Digital twins in healthcare were previously categorized based on the modeling technique used.⁴ Digital twins created using 3D modeling of organs⁴ find applications in tailoring surgical procedures to individual patients.⁵ Digital twin platforms with machine learning as the enabling modeling technique are used to develop clinical decision support systems and generate virtual control patient arms in clinical trials.⁴

Digital twins can also be created using mechanistic simulation models of human physiology,⁹ termed Quantitative Systems Pharmacology (QSP) models. By calibrating a QSP model with patient data, a personalized digital twin closely resembling the patient's physiology can be generated.^{4, 9} Unlike the ML models that learn the system dynamics through patterns in data, QSP models use existing physiological knowledge to simulate causal interactions. This allows clinicians to rationalize and interpret predictions from a QSP model in a clinical setting and helps build their trust in insights derived from CDSS. However, due to the deep domain knowledge requirement, QSP modeling application is limited to diseases with well-understood pathophysiology as compared to ML models.

A hybrid QSP-ML modeling approach addresses limitations in the current understanding of a disease with insights from individual patient data. A digital twin platform for Crohn's Disease (CD)⁹ was developed, based on a hybrid QSP-ML model of CD, to promote shared decision-making in clinical settings. Since frequent endoscopies are impractical, gastroenterologists have limited visibility into the patient's gastrointestinal tract. As a result, they often resort to a “trial and error” method to treat CD patients. The CD digital twin platform simulates endoscopic outcomes in individual patients under different treatment scenarios and helps gastroenterologists understand the impact of treatments on gut tissue damage and mucosal healing, engage patients in conversations (Figure 2) about treatment goals, and jointly develop a treatment plan.⁹

A similar digital twin platform for Type 2 Diabetes (T2D) demonstrated the feasibility of developing personalized nutritional strategies.¹⁰ The T2D digital platform was built on a QSP model of metabolism validated using data from large-scale studies like Diabetes Prevention Program (DPP). These examples demonstrate the value of QSP models beyond their application in drug development. Regardless of the enabling modeling technique, digital twin platforms have tremendous potential in tailoring treatment strategies to individual patients.

Despite AI's numerous advantages in medicine, its adoption remains limited. A 2023 survey by the American Medical Association² found that 65% of physicians see AI's benefits in the medical field. However, 41% were equally excited and concerned about AI in healthcare. Data privacy issues, regulatory hurdles, liability for AI model errors, and a lack of understanding of how AI works² are a few challenges hindering AI's adoption in medicine. Additionally, technical challenges³ in the integration of CDSS such as workflow disruption, interoperability, and alert fatigue also need to be resolved to increase physician adoption of such technologies.

AI development relies on large datasets, but few technical controls exist to inform users about data usage, raising serious data privacy concerns.² Leveraging FAIR data principles (findable, accessible, interoperable, and reusable) can significantly enhance AI model development and validation. Besides data privacy, complex technical issues like model transparency, quality control, data ownership, etc.² need regulation. While the US FDA's Drug Development Tools application and the EMEA's Innovation Task Force offer potential avenues for expedited review, they can initially add considerable time to market. Clear and robust regulations and efficient regulatory pathways are essential for timely AI adoption.

The specter of patient harm from overreliance on AI has fostered physician skepticism. To accelerate AI integration in clinical decision-making, addressing these challenges by cultivating physician trust is paramount. Using QSP models in CDSS, whenever feasible, can help build trust, as physicians can rationalize the CDSS outputs based on the model structure. Trust can also be built by reaffirming that physicians remain ultimate decision-makers and that AI/ML tools provide data-driven insights to aid decision-making. The American Medical Association and American College of Physicians use “Augmented Intelligence”² instead of Artificial Intelligence in clinical practice to emphasize AI's supportive role rather than replacing physicians.

Digital health is driving a cultural transformation in medicine, shifting decision-making from a physician-driven process to shared decision-making by patients and physicians, based on data and analytics.¹ Engaging patients in treatment decisions can improve adherence and overall prognosis. CDSS powered by AI/ML predictive models and QSP models, whether they are diagnostic or prognostic, will play a significant role in augmenting physicians' decision-making abilities, developing optimal treatment strategies, enabling shared decision-making, and thereby revolutionizing healthcare.

No funding was received for this work.

S.P.V., M.G., and H.K. are employees of AstraZeneca and own AstraZeneca stocks or stock options. The authors declared no competing interests for this work.