Machine learning-based body composition analysis for risk stratification in newly diagnosed diffuse large B cell lymphoma

Abstract Frailty is an adverse prognostic factor in cancer, including diffuse large B-cell lymphoma (DLBCL). Although frailty is common among DLBCL patients, there is no consensus on how it can be accurately measured, and it is not included in standard risk assessment. Body composition analysis (BCA) is a promising method for estimating frailty using imaging data. To enable high-throughput BCA based on routine clinical imaging, we developed the Body and Organ Analysis (BOA) pipeline, which allows high-fidelity BCA from computed tomography (CT) data acquired during routine clinical care (Haubold et al., Invest Radiol, 2024). Using BOA, the prognostic impact of body composition in various solid tumors has been demonstrated (Keyl et al., Nat Cancer, 2025). The consensus method for risk assessment in DLBCL is the International Prognostic Index (IPI), which relies on easily accessible clinical features. Although widely accepted, the IPI remains limited in accurately identifying both high-risk and low-risk DLBCL patients. Here, we assessed the utility of BCA using the BOA pipeline to enhance risk stratification in newly diagnosed DLBCL. We computed body composition data for patients from the phase 3 PETAL trial, which evaluated interim PET (iPET) imaging for risk assessment (Dührsen et al., J Clin Oncol, 2018), based on the CT component of PET/CT scans. Patients with a confirmed diagnosis of DLBCL and available imaging data from both initial diagnosis and interim staging were included (n = 291). The median age was 62 years (range: 18–80), and 133 patients (45%) were female. Total metabolic tumor volume (MTV), determined using the ACCURATE tool (Boellaard, J Nucl Med, 2018), was available for all patients. Using BCA, we calculated a sarcopenia index (SI; skeletal muscle volume normalized to bone volume) and a visceral fat index (VFI; ratio of visceral adipose tissue volume to subcutaneous adipose tissue volume) for each patient. Univariable analysis revealed significantly shorter overall survival (OS) in patients with baseline SI in the lowest tertile (log-rank p = 0.002). Furthermore, a greater decrease in SI between baseline imaging and iPET - i.e., during the first two cycles of immunochemotherapy - was associated with shorter OS in the entire cohort (log-rank p = 0.0007) and in the iPET-negative subgroup (log-rank p = 0.0053). Baseline VFI showed a significant association with OS in iPET-negative patients (log-rank p = 0.0432), and a similar trend in the overall cohort (log-rank p = 0.0631). In multivariable analysis using Cox proportional hazards regression, we included SI, VFI, IPI, and MTV as covariates. SI, but not VFI, was confirmed as an independent prognostic variable for OS (HR 0.38, 95% CI 0.18 - 0.79, p = 0.01). To further validate the predictive value of BCA, we trained machine learning models to classify patients into high-risk and low-risk groups based on SI, VFI, MTV, IPI, and gender. Eighty percent of patients were used for training, and the remaining 20% for validation. Among iPET-negative patients, a machine learning model was able to identify a low-risk group comprising one-third of patients with a 5-year OS of 100% (baseline BCA only: log-rank p = 0.026; including iPET BCA: p = 0.0198). Feature importance analysis identified MTV as the most informative feature, followed by SI and VFI. Notably, similar stratification accuracy was achieved with a model using only SI, VFI, IPI, and gender (log-rank p = 0.0222), with SI and VFI ranked as the most important features. In contrast, a model using only IPI and gender failed to meaningfully distinguish risk groups. Altogether, our findings establish sarcopenia as an independent prognostic factor in newly diagnosed DLBCL. Combining BCA with clinical parameters enables the identification of a low-risk subgroup with excellent long-term outcomes. In conclusion, machine learning approaches based on data obtained in routine clinical care can support risk stratification in newly diagnosed DLBCL.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.