Abstract Body

What’s in the notes? Early prediction of cancer using free text in routine primary care data

In the Netherlands about 80% of patients, presenting in general practice with symptoms, possibly indicative of cancer are referred to a specialist within two weeks. Previously, symptom analysis studies (Caper Cards) resulted in better understanding of the early detection of cancer by providing a numerical estimation for the prediction of a prevalent cancer. However, the detection of cancer in clinical practice has not sufficiently improved: of all patients with cancer, about 50% are diagnosed only at an advanced stage. General practitioners still diagnose cancer in roughly the same way as was done decades ago.

New roads have to be explored to improve the detection of cancer at an early stage. Promising development s are new bio-markers, emerging genetic panels, screening and innovative imaging techniques. However, these promising perspectives have not materialized yet.

Developments in artificial intelligence (AI), especially Natural Language Processing (NLP), might provide new possibilities. Although new, until now unknown, symptoms are unlikely to be discovered from the Electronic Patient Record (EPR), information residing in the clinical notes along with other structured data in the EPR could provide additional predictive value.

In order to explore this new pathway, we analysed the medical files of general practitioners of about 1.2 million patients originating from four academic general practice networks, spread over the Netherlands. We validated the cancer diagnoses, starting with lung cancer, by the Dutch cancer registration. We then used NLP together with statistical Machine learning prediction models in order to predict the risk of lung cancer 3 months before the referral was made. Today we present the results of this exciting exercise for the first time. We discuss briefly the used methodology and present the predictive performance in terms of discrimination, calibration and accuracy of the predicted probability and the meaning of these results for clinical practice.

On behalf of the AI-DOC team,

Henk van Weert

The AI-DOC team consists of

Ameen Abu-Hanna. Dpt. Medical informatics, Amsterdam UMC

Martijn Schut, Dpt. Medical informatics, Amsterdam UMC

Torec Luik, Dpt. Medical informatics, Amsterdam UMC

Miguel Rios Goana, Dpt. Medical informatics, Amsterdam UMC

Charles Helsper, Julius Centre, University Medical Center Utrecht

Kristel van Asselt, Dpt. General Practice, Amsterdam UMC

Niek de Wit, Julius Centre, University Medical Center Utrecht

Henk van Weert, Dpt. General Practice, Amsterdam UMC