Accurate classification of pulmonary nodules by a combined model of clinical, imaging, and cell-free DNA methylation biomarkers: a model development and external validation study

Background There is an unmet clinical need for accurate non-invasive tests to facilitate the early diagnosis of lung cancer. We propose a combined model of clinical, imaging, and cell-free DNA methylation biomarkers that aims to improve the classification of pulmonary nodules.Methods We conducted a prospective specimen collection and retrospective masked evaluation study. We recruited participants with a solitary pulmonary nodule sized 5-30 mm from 24 hospitals across 20 cities in China. Participants who were aged 18 years or older and had been referred with 5-30 mm non-calcified and solitary pulmonary nodules, including solid nodules, part solid nodules, and pure ground-glass nodules, were included. We developed a combined clinical and imaging biomarkers (CIBM) model by machine learning for the classification of malignant and benign pulmonary nodules in a cohort (n=839) and validated it in two cohorts (n=258 in the first cohort and n=283 in the second cohort). We then integrated the CIBM model with our previously established circulating tumour DNA methylation model (PulmoSeek) to create a new combined model, PulmoSeek Plus (n=258), and verified it in an independent cohort (n=283). The clinical utility of the models was evaluated using decision curve analysis. A low cutoff (0<middle dot>65) for high sensitivity and a high cutoff (0<middle dot>89) for high specificity were applied simultaneously to stratify pulmonary nodules into low-risk, medium-risk, and high-risk groups. The primary outcome was the diagnostic performance of the CIBM, PulmoSeek, and PulmoSeek Plus models. Participants in this study were drawn from two prospective clinical studies that were registered (NCT03181490 and NCT03651986), the first of which was completed, and the second of which is ongoing because 25% of participants have not yet finished the required 3-year follow-up.Findings We recruited a total of 1380 participants. 1097 participants were enrolled from July 7, 2017, to Feb 12, 2019; 839 participants were used for the CIBM model training set, and the rest (n=258) for the first CIBM validation set and the PulmoSeek Plus training set. 283 participants were enrolled from Oct 26, 2018, to March 20, 2020, as an independent validation set for the PulmoSeek Plus model and the second validation set for the CIBM model. The CIBM model validation cohorts had area under the curves (AUCs) of 0<middle dot>85 (95% CI 0<middle dot>80-0<middle dot>89) and 0<middle dot>85 (0<middle dot>81-0<middle dot>89). The PulmoSeek Plus model had better discrimination capacity compared with the CIBM and PulmoSeek models with an increase of 0<middle dot>05 in AUC (PulmoSeek Plus vs CIBM, 95% CI 0<middle dot>022-0<middle dot>087, p=0<middle dot>001; and PulmoSeek Plus vs PulmoSeek, 0<middle dot>018-0<middle dot>083, p=0<middle dot>002). The overall sensitivity of the PulmoSeek Plus model was 0<middle dot>98 (0<middle dot>97-0<middle dot>99) at a fixed specificity of 0<middle dot>50 for ruling out lung cancer. A high sensitivity of 0<middle dot>98 (0<middle dot>96-0<middle dot>99) was maintained in early-stage lung cancer (stages 0 and I) and 0<middle dot>99 (0<middle dot>96-1<middle dot>00) in 5-10 mm nodules. The decision curve showed that if an invasive intervention, such as surgical resection or biopsy, was deemed necessary at more than the risk threshold score of 0<middle dot>54, the PulmoSeek Plus model would provide a standardised net benefit of 82<middle dot>38% (76<middle dot>06-86<middle dot>79%), equivalent to correctly identifying approximately 83 of 100 people with lung cancer. Using the PulmoSeek Plus model to classify pulmonary nodules with two cutoffs (0<middle dot>65 and 0<middle dot>89) would have reduced 89% (105/118) of unnecessary surgeries and 73% (308/423) of delayed treatments.Interpretation The PulmoSeek Plus Model combining clinical, imaging, and cell-free DNA methylation biomarkers aids the early diagnosis of pulmonary nodules, with potential application in clinical decision making for the of nodules.