Bringing AI to the Clinic for Pancreatic Cancer Care: The Imperative for Seamless Integration into Clinical Workflows

Introduction

In the evolving landscape of precision oncology, the integration of advanced technologies such as artificial intelligence (AI) and machine learning (ML) holds the potential to revolutionize cancer care. Pancreatic cancer, with its notoriously poor prognosis and late-stage detection, stands to benefit significantly from these advancements. Current challenges in pancreatic cancer care include late-stage detection and limited treatment options, making the need for innovative solutions urgent.

AI is poised to revolutionize medicine by interpreting radiographs, ultrasounds, CT, and MRI scans, either complementing or replacing clinician interpretations. Health care organizations can leverage AI to analyze electronic health records (EHRs) for resource allocation, risk identification, and medication error prevention. Using sensors in everyday devices like smartphones and wearables, individuals can share health data to create comprehensive personal health records. Clinicians, in turn, can receive real-time differential diagnoses from AI, enhancing patient care.

Despite its potential, AI adoption in clinical medicine remains limited. The rapid integration of AI into clinical practice hinges on addressing known and unknown challenges and leveraging the vast amounts of digital health data and computational power available. This commentary explores our perspective on how AI and ML can be seamlessly incorporated into clinician workflows to enhance early prediction, detection, and assessment of pancreatic cancer, ultimately improving patient outcomes and addressing the “AI Chasm” (Fig. 1).^1–3 Tools that are created without clinical workflows in mind will face major challenges in both integration and adoption.

OPEN IN VIEWER

Fig. 1.

Leveraging artificial intelligence (AI) for pancreatic cancer will require seamless integration of new tools for early prediction, early detection, and advanced imaging techniques, and will need to be available at the point of care during clinical workflows. The AI chasm will need to assessed with real world evidence and more evidence will be needed before widespread adoption or regulatory changes. PanCan, pancreatic cancer.

Early Prediction Using EHRs

EHRs contain a wealth of patient data that, when analyzed using AI algorithms, can identify patterns and risk factors indicative of early pancreatic cancer. AI can sift through vast amounts of EHR data, including genetic information, family history, lifestyle factors, and clinical symptoms, to identify patients at high risk. Predictive models can then be developed to flag these patients for further diagnostic testing. This proactive approach can lead to earlier intervention and improved survival rates.

A recent study used AI to analyze clinical data from millions of patients in Denmark and the US to predict pancreatic cancer. ⁴ This study was designed to explicitly use of the time sequence of disease events from International Classification of Diseases diagnostic codes and to assess the ability to predict cancer risk for increasing intervals between the end of the disease trajectory used for risk prediction and cancer occurrence The AI models, trained on patients’ medical histories, showed high accuracy in predicting cancer up to 36 months before diagnosis. The best model had an accuracy (AUROC) of 0.88 for Danish data and 0.78 for US data after retraining. These AUROC scores indicate the model’s effectiveness in distinguishing between patients who will and will not develop pancreatic cancer. This approach is particularly promising as it may help identify high-risk patients earlier within the EHR system based on a time sequence of diagnosis codes, allowing for timely intervention and potentially improving survival rates.

Early Detection on Imaging

Early detection of pancreatic cancer is crucial for improving prognosis, and AI integration into medical imaging holds significant promise for enhancing the early detection of pancreatic cancer. Deep learning algorithms can analyze large datasets to produce clearer and more accurate images than traditional methods, aiding in the identification of early-stage abnormalities from CT scans, MRI, and endoscopic ultrasound with high accuracy. ⁵ AI and ML can significantly enhance imaging techniques, enabling the identification of subtle abnormalities that might be missed by the human eye. Techniques like model-based deep learning and algorithm unrolling enhance the interpretability and performance of imaging processes, leading to better feature separation and efficient segmentation. These advancements are crucial for detecting small pancreatic tumors early, which is vital given the aggressive nature of the disease.

Moreover, AI may streamline routine imaging tasks, allowing radiologists to focus on more complex aspects of patient care, such as detailed medical judgments and communication and AI algorithm output review. This shift improves the efficiency of the diagnostic process and supports more precise and timely interventions.

Predictive Analytics of Response for Personalized Medicine

An important study presented at the 2024 ASCO Annual Meeting further exemplifies the potential of AI in personalized cancer treatment. This study, titled “Development and validation of AI-assisted transcriptomic signatures to personalize adjuvant chemotherapy in patients with resectable pancreatic ductal adenocarcinoma,” ⁶ demonstrated the development of transcriptomic signatures that can personalize treatments by predicting patient sensitivity to commonly used pancreatic cancer drugs. The study validated these signatures in a large patient cohort, showing significant improvements in cancer-specific survival and disease-free survival when patients received matched therapies. These findings underscore the potential of AI in tailoring treatment plans to individual patient profiles, enhancing therapeutic outcomes, and reducing unnecessary treatment-related toxicity.

Integration into Clinician Workflows

For AI and ML technologies to be truly transformative, they must be seamlessly integrated into clinician workflows rather than as separate, non-interoperable software solutions that are not readily available at the point of care. Using sensors on common devices like smartphones, wearables, smart speakers, laptops, and tablets, individuals can continuously share health data, contributing to a comprehensive personal health record. These data need to be integrated into the EHR, allowing clinicians to access real-time information. During patient encounters, this integrated data can provide clinicians with real-time differential diagnoses, including probabilities, enhancing the accuracy and efficiency of medical care.

Such a workflow will require the development of user-friendly interfaces and decision support systems that provide real-time insights without overwhelming health care providers. AI tools will need to be interoperable amongst a wide variety of software solutions that hospitals use. The FDA’s Digital Health Innovation Action Plan and the precertification program emphasize the importance of interoperability, and financial incentives could further this goal by rewarding AI devices that demonstrate broad applicability and interoperability.^7,8 Training and education are also critical to ensure that clinicians are comfortable and proficient in using these advanced tools. Collaboration between AI developers, radiologists, oncologists, surgeons and other healthcare professionals is essential to create systems that enhance or evolve, rather than disrupt, clinical practice.

Regulatory and Adoption Challenges

Despite its transformative potential, the adoption of AI in clinical medicine remains limited, primarily confined to modular and niche applications such as image classification tasks that are well-suited to deep learning algorithms. For example, FDA-cleared AI technologies include tools for diagnosing diabetic retinopathy and detecting atrial fibrillation via devices like the Apple Watch. However, clinical impact data post-clearance remains sparse. To date, only AI devices used for assessing coronary artery disease and diagnosing diabetic retinopathy have amassed more than 10,000 CPT claims.^7,8

One of the major barriers to the widespread adoption of AI in clinical settings is the “AI Chasm,” a term that describes the gap between research accuracy and clinical utility. This concept underscores that the accuracy metrics, often measured by the Area Under the Curve (AUC) in research studies, do not necessarily translate to practical clinical utility. AI tools are most effective when trained on data that closely resemble real-world clinical settings, but variability in real-world data can lead to significant errors.

Algorithm bias is a critical issue that can exacerbate health disparities, particularly affecting underrepresented groups. For instance, an AI system trained primarily on data from one demographic may perform poorly on another, as seen in dermatology applications where algorithms have misclassified skin lesions in darker skin tones due to a lack of diverse training data.^9,10 Another notable example is the use of pulse oximeters, which have been shown to be less accurate in individuals with darker skin, highlighting the consequences of biased training data. ¹¹ The “black box” nature of deep learning algorithms fosters distrust among clinicians and patients, as these algorithms often lack transparency. For example, a study found that an AI model designed to classify chest X-rays inadvertently used the presence of a radiologist’s notes as a proxy for disease, leading to misclassifications. This lack of interpretability makes it difficult to identify and correct such issues.

However, strategies exist to mitigate bias and ensure equity. AI models that are trained on diverse datasets that represent various demographics can help reduce bias. Collaborating with international institutions and using federated learning can enhance data diversity without compromising patient privacy. Regular bias audits using fairness metrics can help identify and address disparities in AI model performance. Metrics such as demographic parity, equalized odds, and disparate impact can be used to evaluate model fairness. Furthermore, developing explainable AI techniques can enhance transparency and trust. Model-agnostic methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) can help clinicians understand AI decision-making processes and identify potential biases. ¹² Finally, engaging ethicists, sociologists, and patient advocacy groups in the development and deployment of AI systems can ensure that diverse perspectives are considered, and potential biases are addressed proactively. The scarcity of prospective studies assessing the efficacy of AI in real-world settings further contributes to hesitancy in adoption. Unlike traditional medical devices, AI systems face unique regulatory challenges, including continuous learning and adaptation of algorithms, which complicate the approval process.

Overcoming these obstacles requires significant investment and expertise, similar to the pharmaceutical industry’s approach, which anticipates multiple failures before achieving success. Evolving regulatory frameworks aim to balance innovation with safety and effectiveness, ensuring equitable benefits across populations. Despite these evolving processes, payment mechanisms for AI-based medical devices continue to lag behind and evolve slowly.

Addressing these challenges necessitates increased publication of peer-reviewed studies in diverse clinical settings to build robust real-world evidence of AI efficacy. There is also a need for regulatory processes that are adaptive and responsive to the unique challenges posed by AI technologies. Large-scale investment in AI research and development, coupled with interdisciplinary collaboration, will be crucial in bridging the AI Chasm and realizing the full potential of AI in clinical medicine.

Education and Training for Health care Professionals

Another critical component for the successful adoption of AI in clinical practice is the education and training of health care professionals. Effective strategies have only begun to be developed, but could include:

Comprehensive Training Programs: Implementing structured training programs that cover the fundamentals of AI, its applications in health care, and specific tools relevant to the clinicians’ practice areas. These programs should include both theoretical knowledge and practical hands-on sessions. ¹³

Conclusion and Future Directions

The integration of AI and ML into pancreatic cancer care promises significant improvements in early prediction, detection, and staging, ultimately enhancing patient outcomes. However, realizing this potential requires addressing several challenges. Data privacy and security must be prioritized to protect patient information. Ensuring the accuracy and reliability of AI algorithms is critical to prevent misdiagnoses. Regulatory frameworks need to evolve to keep pace with rapid technological advancements, supporting the safe and effective implementation of AI in clinical settings.

Ongoing research and clinical trials are essential to validate the effectiveness of AI applications in pancreatic cancer. Interoperability is crucial for AI tools to be effective across different health systems and patient populations. Incentivizing the development of generalizable and unbiased AI technologies through regulatory and payment frameworks can further this goal. For instance, providing greater reimbursements for AI devices that demonstrate broad applicability and mitigate bias during premarket testing can motivate developers to create more inclusive and reliable tools.

By integrating AI seamlessly into clinical workflows and ensuring these technologies are properly incentivized and regulated, we can make significant strides in the fight against pancreatic cancer. Embracing these advancements will enable personalized and effective care, ultimately improving patient outcomes and transforming the landscape of precision oncology.