Beyond Numbers: Making AI Work for Patients with Advanced Cancer

The Evolving Challenge of Cancer Prognostication

Over the last decade, I've witnessed a remarkable transformation in my supportive oncology practice. The advent of targeted therapies and immunotherapy has moved us beyond the traditional triad of surgery, chemotherapy, and radiation. While these advances have dramatically improved the survival of many patients, they've paradoxically made one of our core responsibilities—prognostication—more complex than ever.

Understanding Modern Cancer Trajectories

 What was once a relatively predictable pattern of gradual decline (in advanced cancer) has evolved into multiple distinct paths, each reflecting modern cancer care's complexity:

 - Dramatic improvements with immunotherapy (yellow line), as I've seen in melanoma patients who've returned to work and travel

- Extended stability with targeted therapies (lavender line), exemplified by some lung cancer patients who continue working full-time years after diagnosis

- Traditional gradual decline (blue line), still typical in many solid tumors

- Sudden deterioration from treatment complications (red line), a stark reminder of our treatments' double-edged nature

These aren't just abstract patterns on a mortality curve – they represent real people making crucial decisions about their lives.

AI's Emerging Role in Prognostication

 Researchers have turned to artificial intelligence to seek solutions to this prognostic complexity. At Duke University, a study implemented a machine learning algorithm to identify high-mortality-risk patients during oncology unit admissions. The system automatically alerted clinical teams when it predicted poor prognosis, dramatically increasing goals-of-care discussions among high-risk patients from 2% to 80%.

The Timing Paradox

 However, these impressive numbers tell only part of the story. Despite more frequent discussions, the study showed no significant changes in crucial end-of-life metrics:

- length of hospital stay

- inpatient mortality

- ICU admission rates

- 30-day readmissions

 There are many potential explanations for this disconnect, but one stands out: timing matters profoundly. Too often, these conversations happen during moments of crisis in the hospital, when patients and families are overwhelmed. As another patient reflected, "If wished we'd had these discussions earlier."

Beyond Prediction: The Human Element

 While AI can excel at pattern recognition and risk prediction, it will not replace the nuanced art of prognostic communication. As medical ethicist Sarah C. Hull notes, "AI can't worry about patients."

Moving Forward Together

 The promise of AI in cancer prognostication lies in helping us identify opportunities for meaningful conversations earlier in the disease trajectory when patients are still well enough to participate fully in decision-making about their care. The goal isn't just to predict outcomes but to help our patients live as fully as possible, aligned with their values and wishes.

These tools could help us move from crisis-driven discussions to more timely, thoughtful conversations about the future. In doing so, we might better serve our patients in navigating the increasingly complex landscape of modern cancer care.

Further Reading:

• Geijteman, E. C., Kuip, E. J., Oskam, J., Lees, D., & Bruera, E. (2024). Illness trajectories of incurable solid cancers. bmj, 384.

• Patel, M. N., Mara, A., Acker, Y., Gollon, J., Setji, N., Walter, J., ... & Ma, J. (2024). Machine Learning for Targeted Advance Care Planning in Cancer Patients: A Quality Improvement Study. Journal of Pain and Symptom Management, 68(6), 539-547.

• Ethics of AI in Medicine”: Hull SC, Interview on AI's role in clinical care. Hswen, Y., & Abbasi, J. (2024). AI Can’t Worry About Patients, and a Clinical Ethicist Says That Matters. JAMA.

#oncology #artificialintelligence #palliativecare #healthcareinnovation #patientcare