Artificial intelligence for early detection of lung cancer in GPs' clinical notes: a retrospective observational cohort study

Abstract

Background: The journey of >80% of patients diagnosed with lung cancer starts in general practice. About 75% of patients are diagnosed when it is at an advanced stage (3 or 4), leading to >80% mortality within 1 year at present. The long-term data in GP records might contain hidden information that could be used for earlier case finding of patients with cancer.

Aim: To develop new prediction tools that improve the risk assessment for lung cancer.

Design and setting: Text analysis of electronic patient data using natural language processing and machine learning in the general practice files of four networks in the Netherlands.

Method: Files of 525 526 patients were analysed, of whom 2386 were diagnosed with lung cancer. Diagnoses were validated by using the Dutch cancer registry, and both structured and free-text data were used to predict the diagnosis of lung cancer 5 months before diagnosis (4 months before referral).

Results: The algorithm could facilitate earlier detection of lung cancer using routine general practice data. Discrimination, calibration, sensitivity, and specificity were established under various cut-off points of the prediction 5 months before diagnosis. Internal validation of the best model demonstrated an area under the curve of 0.88 (95% confidence interval [CI] = 0.86 to 0.89), which shrunk to 0.79 (95% CI = 0.78 to 0.80) during external validation. The desired sensitivity determines the number of patients to be referred to detect one patient with lung cancer.

Conclusion: Artificial intelligence-based support enables earlier detection of lung cancer in general practice using readily available text in the patient files of GPs, but needs additional prospective clinical evaluation.

Keywords: early detection; general practice; lung cancer; machine learning; natural language processing; oncology.

Figure 1.

Precision-recall curves of the phrase skip-gram (PSG) model when used with text only (TO) and when used with text and coded data (TC). PPV = positive predictive value. PSGNN = phrase skip-gram neural network.

Figure 2.

Calibration plot of the phrase skip-gram (PSG) model when used with text only (TO) and when used with text and coded data (TC). The plot shows actual probabilities (y-axis) versus predicted probabilities (x-axis), and an ideal curve (dotted line) is included for illustrating when the predicted probabilities are identical to the actual probabilities. PSGNN = phrase skip-gram neural network.