Using a combined approach of Kernel machine-learning (ML) and random optimization (RO) techniques we recently developed a set of predictors (ML-ROs) for VTE risk assessment. Aim of this study was to validate a model incorporating the two best ML-ROs to devise a web-based graphical interface for VTE risk stratification.
Pre-chemotherapy age, sex, tumor site and stage, hematological attributes, fasting blood lipids, glycemic indexes, liver and kidney function, BMI, ECOG, supportive and anti-cancer drugs of 608 cancer outpatients were entered in the model, with numerical attributes analyzed as continuous values. Variables were clustered into groups according to clinical significance, and RO was used to devise their relative weight in final prediction.
VTE occurred in 7.1% of patients. Overall, 6% were at high-risk for VTE, as per current guidelines (Khorana Score (KS) ≥3), 11% of which had VTE during treatment. 42% and 52% were at intermediate (KS 1/2) or low-risk (KS = 0), with VTE rates of 9% and 5%, respectively. Accordingly, the performance of KS, despite a 94% specificity, was characterized by a 9% sensitivity with an area under the ROC curve (AUROC) of 0.589, translating into non-significant positive (+LR) [1.58 (0.48-4.30)] or negative likelihood ratio (-LR) [0.96 (0.83-1.04)]. Conversely, the VTE risk prediction performance of the combined ML model showed a 0.716 AUROC, which was significantly higher than that observed with KS (difference between areas: 0.127, p = 0.0044). At a criterion >1 (risk estimate achieved by both predictors) this combined approach showed significant +LR [2.30 (1.70-2.82)] and -LR [0.46 (0.28-0.69)] and a 4.9 Hazard Ratio (95%CI: 2.5-9.4) with a 6-month VTE rate of 3.4% in the low-risk, compared with 14.9% in the high-risk category.
These results demonstrate that a ML approach, optimizing the relative weight (by RO) of groups of clinical attributes, is of clinical value for VTE risk prediction, performing better than KS. We are now finalizing the architecture of a web service with a graphical interface helping oncologists in the critical phase of decision making.
Clinical trial identification
Legal entity responsible for the study
RISK Research Group
European Social Fund PON03PE_00146_1/10 BIBIOFAR
All authors have declared no conflicts of interest.