1171MO - AI-derived diagnostic criteria for early lung cancer screening using primary care health records

Background

Detecting lung cancer in its early stages poses a significant challenge due to the lack of specific diagnostic markers. The digitisation of healthcare records has generated vast amounts of untapped data, presenting opportunities for improvements in early diagnoses. This study presents a novel approach to mine and extract data to automatically form diagnostic criteria for lung cancer from primary care appointment data.

Methods

Patients aged 40 years or older, referred for a chest x-ray between 2016 and 2022, were included. Lung cancer status was determined using the primary and secondary care electronic health records, and the Somerset Cancer Registry. Patients with pre-existing cancer diagnoses were excluded. Coded diagnoses and symptoms were extracted from primary care records to establish diagnostic criteria. A custom feature selection algorithm was used to identify the most relevant historical diagnoses and symptoms, whilst minimising redundancy among predictors. The selected features were used to train a Random Forest classifier using 75% of the training data, validating on the remaining 25%. The resulting model was applied longitudinally to a test set of patients to determine the first time a patient is flagged for screening.

Results

The diagnostic model was trained on data from 78,457 patients, using age, sex, ethnicity, and the presence of any of the following predictors, within 6 months of the appointment: bronchiolar disease, chronic lung disease, dyspnoea, ex-smoker, cough and altered bowel function. An AUROC of 0.80 was achieved, with a sensitivity and specificity of 87% and 57%, respectively. When longitudinally applied to the test set, the model would have detected 77% of lung cancer patients, offering a median “advanced notice” of 36 months before diagnosis (IQR = 19 - 52 months).

Conclusions

This study demonstrated success in leveraging primary care data to establish diagnostic criteria for lung cancer. The performance of the model to identify lung cancer years earlier requires further validation, but such an approach demonstrates promise for enhancing early detection. Future efforts will concentrate on prospectively evaluating the developed criteria and determining their health and economic value.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Barts Health NHS Trust.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.