Machine learning-based survival prediction models for advanced-stage Hodgkin lymphoma

Accurate prognostication is important for the optimal treatment selection in patients with classical Hodgkin lymphoma.¹ The International Prognostic Score (IPS)-7 is typically used to risk-stratify patients with advanced-stage Hodgkin lymphoma (aHL).¹ The IPS-3 model was developed to simplify the IPS-7 while retaining its predictive capabilities. Recently, the aHL International Prognostic Index (A-HIPI) was developed to improve prognostication compared with the IPS models.¹

New predictive models developed using machine learning (ML) have the potential to further improve prognostication. Jørgensen et al.¹ recently developed and validated an ML-based survival prediction model for patients with aHL and published the findings in Journal of Clinical Oncology Clinical Cancer Informatics. Below, we summarize the key points.

Methods and patient population¹

• ML models were developed using data from adult patients with newly diagnosed aHL from the Danish National Lymphoma Register (development cohort).
• The model was validated using data from adult patients with newly diagnosed aHL from the Cancer Registry of Norway and the Swedish Lymphoma Register (validation cohort).
• Performance was measured by the integrated Brier score, the time-varying area under the curve (AUC), and the concordance index (C-index).

Key findings¹

• In total, 1,467 patients were included (development cohort, n = 707; validation cohort, n = 760).
• Median follow-up was 7.2 years and 8.6 years in the development and validation cohorts, respectively.
• The ML model achieved the highest C-index for overall survival (OS) and progression-free survival (PFS) in both the development and validation cohorts (Table 1).

Table 1. Performance of the ML, IPS-3, IPS-7, and A-HIPI models in the development and validation cohorts*

Model	Development cohort (n = 707)				Validation cohort (n = 760)
OS		PFS		OS		PFS
IBS	C-index	IBS	C-index	IBS	C-index	IBS	C-index
ML	0.073	0.789	0.141	0.665	0.049	0.749	0.085	0.691
IPS-3	0.085	0.650	0.149	0.576	0.054	0.663	0.078	0.64
IPS-7	0.087	0.608	0.151	0.549	0.052	0.700	0.077	0.672
A-HIPI	0.076	0.768	0.142	0.649	0.052	0.741	0.091	0.677
A-HIPI, advanced-stage Hodgkin lymphoma international prognostic index; C-index, concordance index; IBS, integrated Brier score; IPS, International Prognostic Score; ML, machine learning; OS, overall survival; PFS, progression-free survival. *Adapted from Jørgensen, et al.1

• The time-varying AUC for OS and PFS was consistently higher for the ML model vs IPS-7 and IPS-3 in both cohorts.
  • The time-varying AUC for the A-HIPI model was similar to the ML model for OS and PFS in the development cohort; this was similar for OS in the validation cohort.
  • However, time-varying AUC for PFS was similar to the IPS-7 in the validation cohort.
• The 5-year OS and PFS rates for patients classified as high-risk by the ML and IPS-7 models are shown in Figure 1.

IPS, International Prognostic Score; ML, machine learning; OS, overall survival; PFS, progression-free survival.
*Data from Jørgensen, et al.¹

 

Key learnings
The ML model showed improved prognostication for patients with aHL vs the IPS-7 and IPS-3. A high-risk group was identified with worse OS than high-risk patients under the IPS-7 model. The ML model was only slightly better than the recently developed A-HIPI model.