{
  "abstract": "Introduction Malignancy risk assessment of incidental pulmonary nodules (IPNs) is key to identifying early-stage lung cancer. IDEAL, a multi-centre prospective observational study, assessed the utility of an AI algorithm in patients with solid IPNs.Here we report the overall diagnostic performance of the AI algorithm and compare its performance to the Brock model.Methods Participants aged ≥18 years with fewer than five 5–15 mm solid IPNs identified on CT were prospectively recruited from four NHS hospitals between 08/2018 and 02/2022. Patients with ground-glass opacities and history of malignancy within five years were excluded.Management of nodules was as per the 2015 BTS Pulmonary Nodule Guidelines. Final diagnosis was benign if the nodule was classified as categorically benign on CT requiring no follow-up or has completed surveillance as per BTS; or malignant if proven histologically. The AI provided a CAP score on selected IPNs.Brock score threshold of 10% was compared to a predetermined rule out AI threshold risk of 0.56334234.Results 1056 patients with 1594 nodules were included. 24/1056 (2.27%) patients had histologically-proven malignant nodules with one patient having two synchronous malignancies. Of the 25/1594 (1.57%) malignant nodules, 23 were primary lung cancers; adenocarcinoma was the most common.404 nodules (20 malignant) were selected for analysis: these were solid, ≥8 mm in diameter (as per BTS). The Area Under the Curve (AUC) for the algorithm was 79.3% (95%CI 68.54%–90.05%), compared with 79.56% (95%CI 71.01%–88.08%) for Brock (p=0.995).For the algorithm, there were 42 nodules (0 malignant) below and 362 nodules (20 malignant) above the predetermined threshold. For Brock, there were 290 nodules (8 malignant) below and 114 nodules (12 malignant) above the 10% threshold. When comparing at equivalent sensitivities, i.e. ruling out the same number of benign nodules as Brock below 10% (282 nodules), CAP would exclude 6 malignant nodules rather than 8 as with Brock.Abstract S31 Figure 1Scatterplot illustrating the relationship of CAP score and Brock model score of the 404 nodules included in the analysis. The reference lines are pre-defined rule-out thresholdsConclusion This is the first study testing the utility of AI for 5–15 mm IPNs prospectively against current standard of care. Despite a comparable AUC with Brock, AI appears to more reliably identify benign nodules without requiring any clinical information, in comparison to the Brock model.",
  "authors": [
    {
      "affiliations": [
        "University of Oxford, Oxford, UK"
      ],
      "name": "KL Ng"
    },
    {
      "affiliations": [
        "University of Oxford, Oxford, UK"
      ],
      "name": "P Dhiman"
    },
    {
      "affiliations": [
        "Oxford University Hospitals NHS Foundation Trust, Oxford, UK"
      ],
      "name": "JA Moreland"
    },
    {
      "affiliations": [
        "Oxford University Hospitals NHS Foundation Trust, Oxford, UK"
      ],
      "name": "W Hickes"
    },
    {
      "affiliations": [
        "University of Oxford, Oxford, UK"
      ],
      "name": "ML Szkonter"
    },
    {
      "affiliations": [
        "University of Oxford, Oxford, UK"
      ],
      "name": "N Rahman"
    },
    {
      "affiliations": [
        "Royal Berkshire NHS Foundation Trust, Reading, UK"
      ],
      "name": "T Barton"
    },
    {
      "affiliations": [
        "Leeds Teaching Hospital NHS Trust, Leeds, UK"
      ],
      "name": "M Callister"
    },
    {
      "affiliations": [
        "Nottingham University Hospitals NHS Trust, Nottingham, UK"
      ],
      "name": "D Baldwin"
    },
    {
      "affiliations": [
        "University of Oxford, Oxford, UK"
      ],
      "name": "FV Gleeson"
    }
  ],
  "title": "S31 Early insights into the data from the artificial intelligence and big data for early lung cancer diagnosis prospective study (phase 2) (IDEAL study)",
  "uid": "1f5f25d2-a9ac-5644-8b33-74265120db5a"
}