817P - A machine learning algorithm utilizing clinicopathologic parameters for extranodal natural killer/T cell lymphoma

Background

Current clinical prediction indexes lack the requisite accuracy for extranodal natural killer/T-cell lymphoma (ENKTL) following non-anthracycline-based therapy. This study addresses this gap by introducing and validating a machine learning (ML) algorithm based on clinicopathologic parameters to effectively stratify risk among ENKTL patients.

Methods

Data from 977 patients with ENKTL from 3 cohorts of China were analyzed. 15 ML algorithms were trained and validated for overall survival (OS) prediction in 644 patients from Sun Yat-sen University Cancer Center. The optimization of hyperparameters was conducted through an exhaustive ten-fold cross-validation grid search. External validations were performed in two external cohorts (N=187 from West China Hospital, and N=146 from Sichuan Cancer Hospital, respectively). Evaluation of model performance included Harrell’s c-index, receiver operating characteristic (ROC) curve, calibration curve, and decision curve analysis.

Results

Among the 15 ML algorithms tested, Gradient Boosting Machine (GBM) demonstrated superior performance. The ENKTL-ML score, generated by GBM, incorporated 13 clinicopathologic parameters. In external validation cohorts, the c-indexes for the ENKTL-ML score were 0.84 (95% confidence interval [CI]: 0.81, 0.88) and 0.83 (0.72, 0.94), respectively. The area under ROC curves for 5-year OS were ≥ 0.81. The ENKTL-ML score outperformed existing indexes in discriminatory ability and treatment recommendations. Notably, it effectively stratified patients into low-, intermediate-, and high-risk groups with distinctive survival outcomes. The ENKTL-ML score demonstrated the ability to differentiate low-risk patients from those at stage I. Adding anti-PD-1 immunotherapy as first-line treatment significantly improved survival in the high-risk group, with no substantial benefit observed in low/intermediate-risk patients. An online calculator for the ENKTL-ML score is accessible online.

Conclusions

The ENKTL-ML score serves as an innovative machine learning tool, facilitating risk stratification and aiding clinicians in decision-making for ENKTL patient management.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Y. Huang.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.