Machine learning to identify a composite indicator to predict cardiac death in ischemic heart disease.

BACKGROUND: Machine learning (ML) employs algorithms that learn from data, building models with the potential to predict events by aggregating a large number of variables and assessing their complex interactions. The aim of this study is to assess ML potential in identifying patients with ischemic heart disease (IHD) at high risk of cardiac death (CD). METHODS: 3987 (mean age 68 ± 11) hospitalized IHD patients were enrolled. We implemented and compared various ML models and their combination into ensembles. Model output constitutes a new ML indicator to be employed for stratification. Primary variable importance was assessed with ablation tests. RESULTS: An ensemble classifier combining three ML models achieved the best performance to predict CD (AUROC of 0.830, F1-macro of 0.726). ML indicator use through Cox survival analysis outperformed the 18 variables individually, producing a better stratification compared to standard multivariate analysis (improvement of ∼20%). Patients in the low risk group defined through ML indicator had a significantly higher survival (88.8% versus 29.1%). The main variables identified were Dyslipidemia, LVEF, Previous CABG, Diabetes, Previous Myocardial Infarction, Smoke, Documented resting or exertional ischemia, with an AUROC of 0.791 and an F1-score of 0.674, lower than that of 18 variables. Both code and clinical data are freely available with this article. CONCLUSION: ML may allow a faster, low-cost and reliable evaluation of IHD patient prognosis by inclusion of more predictors and identification of those more significant, improving outcome prediction towards the development of precision medicine in this clinical field.

HOXD8 hypermethylation as a fully sensitive and specific biomarker for biliary tract cancer detectable in tissue and bile samples.

BACKGROUND: Biliary tract cancers (BTC) are rare but highly aggressive tumours with poor prognosis, usually detected at advanced stages. Herein, we aimed at identifying BTC-specific DNA methylation alterations. METHODS: Study design included statistical power and sample size estimation. A genome-wide methylation study of an explorative cohort (50 BTC and ten matched non-tumoral tissue samples) has been performed. BTC-specific altered CpG islands were validated in over 180 samples (174 BTCs and 13 non-tumoral controls). The final biomarkers, selected by a machine-learning approach, were validated in independent tissue (18 BTCs, 14 matched non-tumoral samples) and bile (24 BTCs, five non-tumoral samples) replication series, using droplet digital PCR. RESULTS: We identified and successfully validated BTC-specific DNA methylation alterations in over 200 BTC samples. The two-biomarker panel, selected by an in-house algorithm, showed an AUC > 0.97. The best-performing biomarker (chr2:176993479-176995557), associated with HOXD8, a pivotal gene in cancer-related pathways, achieved 100% sensitivity and specificity in a new series of tissue and bile samples. CONCLUSIONS: We identified a novel fully efficient BTC biomarker, associated with HOXD8 gene, detectable both in tissue and bile by a standardised assay ready-to-use in clinical trials also including samples from non-invasive matrices.