The Somerset NHS Foundation Trust deployed the red dot®CXR solution for a three-month pilot period to evaluate the use of the algorithm in supporting the Trust meet its national targets for lung cancer diagnosis.

To monitor the accuracy of the red dot®CXR algorithm, radiologists indicated whether they agreed or disagreed with the algorithm’s classification in their exam text reports. All exams reported as disagreements were reviewed and arbitrated by experienced radiologists. To maximize patient safety and performance standards throughout the trial, all exams classified as HCN by the red dot®CXR solution were still reported by a radiologist.

Of the 3,794 images processed during the trial, the model classified 147 (3.9%) exams as SLC, and 562 (14.8%) as HCN. On average, the red dot® algorithm returned a result in 15.7 seconds.

Within this trial, Behold.ai demonstrated a notable reduction in the time taken to progress patients through the lung cancer pathway, and it showed that the real performance of our algorithm closely matches that of our peer-reviewed published results [1]. In more detail, the red dot® algorithm demonstrated time-saving benefits for the SomersetNHS Foundation Trust that can promote faster diagnosis for patients, as the deployment of the algorithm more than halved the time between initial X-ray screening and a CT scan (the gold standard for detecting lung cancer), from 7 days to 3 days.

The trial demonstrated the effectiveness of red dot® platform for flagging exams as SLC. As previously mentioned, the model flagged 147 (3.9%) exams as Suspected Lung Cancer (SLC), 19 of which (12.9%) were later diagnosed with cancer. Of the exams that were classed as SLC by the model, 75% were reported within 6 hours by the Trust. These figures indicate that scans with suspected lung cancer are being effectively prioritised for reporting and fast-tracking diagnosis.

The algorithm ruled out 14.8% of total scans as High Confidence Normal (HCN), effectively indicating that they can be removed from radiologists’ workload and auto-reported. It did so with a high degree of accuracy; the model performed with a negative predictive value (NPV) of 97.7% and a false negative rate of 0.34%. This saves time and resources by easing the burden on radiologists and allowing them to prioritise their workload. The few discrepancies (13 out of 562 cases) that were found between hospital radiologists and the model’s HCN classification were arbitrated by two independent hospital radiologists, and none of these discrepancy cases were found to have clinically significant implications.