How Can AI Improve Lung Cancer Detection Directly in GP Clinics?

An Algorithm That Reads Between the Lines

The model was trained on data from over 1.1 million patients whose medical records span up to 30 years. It works with complete medical documentation and combines structured data (age, gender, lab results) with unstructured texts, such as seemingly minor notes taken by doctors during patient visits.

"Our model can extract predictive signals from medical history data that are often overlooked. This allows us to detect the risk of lung cancer up to four months in advance," says Professor Martijn Schut, a translational AI expert from Amsterdam UMC.

Previously Overlooked Records Now Speak Clearly

Unlike studies focused only on obvious factors like smoking or hemoptysis, the model also reads between the lines. It uncovers subtle signals hidden in doctors' notes – details that could otherwise be easily missed. It learns to recognize patterns from inconspicuous symptoms like fatigue, coughing, weight loss, or shortness of breath – signs that may appear in the documentation months before a diagnosis.

The algorithm uses advanced variable selection methods (e.g., Akaike information criterion) to identify specific patterns in order to estimate cancer probability as accurately as possible. "This model allows us to recognize warning signs within the context of individual medical histories," explains Professor Ameen Abu Hanna, a clinical informatics specialist.

However, exactly which linguistic fragments play a key role still requires further research. This will be critical for deploying the system in everyday practice.

Fewer False Alarms, Greater Chances of Survival

Unlike some screening tools, the new approach generates significantly fewer false positives. Risk assessment takes place without added burden – during a routine GP visit.

"Most lung cancer patients arrive late, often at stage 3 or 4. This reflects in a poor prognosis – about 80% die within one year of diagnosis," reminds Henk van Weert, emeritus professor of general practice.

Lung cancer remains one of the most common and serious cancer diagnoses. Despite advances in therapy, the five-year mortality rate remains over 80%. According to the emeritus professor, even an earlier diagnosis by just four weeks could significantly affect survival rates.

Promising Results from Real-World Practice

The study drew data from 525,526 patients in GP clinics in Amsterdam, Utrecht, and Groningen. Lung cancer was diagnosed in 2,386 of them, and the diagnoses were verified using the Netherlands Cancer Registry. AI successfully identified 62% of patients who were later confirmed to have cancer.

The results highlight not only a potential improvement in prognosis but also a possibility to reduce treatment costs through earlier intervention. A similar approach could in the future help detect other cancers that are often identified late – such as pancreatic, stomach, or ovarian cancer.

Challenges for the Future

Implementing the AI model into everyday practice won’t be without obstacles. Further clinical validation is needed across various healthcare systems and settings. Transparency is also an issue – doctors must understand how the model arrives at its recommendations and be able to interpret them. Practitioners’ willingness to adopt new technologies will also be crucial, especially given their already high time and administrative burdens.

Despite these challenges, it is clear that algorithms capable of working with real-world data could represent a major shift in diagnostics – not only in oncology.

Editorial Team, Medscope.pro

Sources:
1. Schut M. C., Luik R. T., Vagliano I. et al. Artificial intelligence for early detection of lung cancer in GPs’ clinical notes: a retrospective observational cohort study. BJGP, doi: 10.3399/BJGP.2023.0489.

2. Lung cancer detected four months earlier at GP with artificial intelligence. Cancer Center Amsterdam. Available at: www.amsterdamumc.org/en/research/institutes/cancer-center-amsterdam/news/lung-cancer-detected-four-months-earlier-at-gp-with-artificial-intelligence.htm?utm_source=chatgpt.com