Comparison of deep learning-based recurrence-free survival with random survival forest and Cox proportional hazard models in Stage-I NSCLC patients

  • Irem Kar ,
  • Gokhan Kocaman ,
  • Farrukh İbrahimov ,
  • Serkan Enön ,
  • ,
  • Atilla Halil Elhan

Cancer Medicine |

DOI

Background
The curative treatment for Stage I non-small cell lung cancer (NSCLC) is surgical resection. Even for Stage I patients, the probability of recurrence after curative treatment is around 20%.

Methods
In this retrospective study, we included 268 operated Stage I NSCLC patients between January 2008 and June 2018 to analyze the prognostic factors (pathological stage, histological type, number of sampled mediastinal lymph node stations, type of resection, SUVmax of the lesion) that may affect relapse with three different methods, Cox proportional hazard (CoxPH), random survival forest (RSF), DeepSurv, and to compare the performance of these methods with Harrell’s C-index. The dataset was randomly split into two sets, training and test sets.

Results
In the training set, DeepSurv showed the best performance among the three models, the C-index of the training set was 0.832, followed by RSF (0.675) and CoxPH (0.672). In the test set, RSF showed the best performance among the three models, followed by DeepSurv with 0.677 and CoxPH methods with 0.625.

Conclusion
In conclusion, machine-learning techniques can be useful in predicting recurrence for lung cancer and guide clinicians both in choosing the adjuvant treatment options and best follow-up programs.