Predicting Infection Risk in CLL Using Domain Adaptation

Domain adaptation may be a novel creative solution to predict infection risk in patients with chronic lymphocytic leukemia (CLL), infection being a leading cause of reduced quality of life and increased health care costs in these patients because of treatment-related immune suppression.^1,2

The results of a new analysis published in Acta Oncologica show that better risk prediction tools are needed—with clinicians currently assessing infection risk using basic factors like age and comorbidities—but also that that building them is difficult. CLL is relatively rare, and machine learning models require large amounts of data to perform well.³ Domain adaptation trains a model on data from a related disease to compensate for limited data in the target condition. The investigators wanted to test whether records from patients with lymphoma could be leveraged to improve infection prediction in CLL.

Researchers at Copenhagen University Hospital drew on the DALY-CARE data resource, a Danish registry linking electronic health records from patients with hematologic cancers across eastern Denmark, extracting such data as medication history, biochemistry results, prior diagnoses, and hospitalization records. The study encompassed 301 patients with CLL and 2397 patients with lymphoma who received first-line treatment between May of 2016 and August of 2023. Median (IQR) ages at diagnosis were 68 years (60-75) and 68 years (56-76), respectively.

The investigators defined severe infection as receiving intravenous antimicrobials and a concurrent blood culture draw, within 2 days of one another, within 1 year of treatment initiation. Three prediction strategies were compared: a domain-specific model (CLL data only), a domain adaptation model (trained on both CLL and lymphoma data, tested on CLL only), and the standard CLL International Prognostic Index (CLL-IPI). Logistic regression, random forest, and gradient boosting were the classifier types tested across the 3 strategies.

Key Findings

The domain adaptation approach outperformed the others across every model tested. Using the best-performing classifier (random forest), the domain adaptation model achieved a Matthews correlation coefficient of 0.342 vs 0.290 for the CLL-only model and 0.166 for the CLL-IPI score alone. In terms of clinical stratification, domain adaptation also separated high-risk patients from low-risk patients with an OR of 4.43 compared with 3.69 for the domain-specific model and 2.27 for CLL-IPI.

Domain adaptation also relied heavily on medication-derived features, including antibacterial drugs, alimentary tract medications, and nontherapeutic products, while the CLL-only model leaned more on biochemistry results like albumin and C-reactive protein (CRP). Elevated CRP was identified as a predictive feature by both approaches, consistent with its role as a marker of infection or systemic inflammation.⁴

Predictive accuracy also was notably stronger for patients with a prior hospitalization history, since those records provide richer data on comorbidities and frailty, the authors explained. For patients without hospitalization history, all models performed worse.

Real-World Implications

The findings carry 2 important messages for oncology and health informatics. First, when disease-specific data are scarce, borrowing from a biologically adjacent condition can meaningfully improve predictive performance. CLL and lymphoma are distinct, but they share enough clinical overlap (eg, similar treatment pathways, overlapping comorbidity profiles, comparable infection dynamics) that lymphoma data have potential to add a genuine signal.

Second, domain adaptation highlights that infection risk in CLL is embedded in the broader clinical picture before treatment begins, as captured in medication lists, lab values, and hospitalization patterns rather than genomic variables alone. This suggests that rich electronic health record data, already routinely collected, could be harnessed to flag high-risk patients before treatment decisions are made.

The authors underscore that their present results echo those from a previous analysis, such as the domain adaptation and specific strategies had better predictive performance among recently hospitalized patients.

“From a modeling perspective,” they conclude, “employing [domain adaptation] techniques that learn the shifts in data across diseases, along with the use of deep learning models, may be a promising approach to explore.”

The authors note that future work incorporating data from multiple myeloma or treatment-naïve patients with CLL, or employing deep learning architectures, could further strengthen the approach. Still, their work builds a credible path toward data-driven infection risk stratification in CLL, an approach that is showing potential to inform treatment selection, prophylactic antibiotic strategies, and monitoring intensity for patients.

Their study has limitations. Their patient cohort was small, and the domain adaptation approach was tested only with lymphoma as the source domain. Also, the model performed poorly for patients without hospitalization histories, leaving a gap for healthier, less data-rich patients.

References

1. Ruppert AS, Booth AM, Ding W, et al. Adverse event burden in older patients with CLL receiving bendamustine plus rituximab or ibrutinib regimens: Alliance A041202. Leukemia. 2021;35(10):2854-2861. doi:10.1038/ S41375-021-01342-X
2. Goyal RK, Nagar SP, Kabadi SM, Le H, Davis KL, Kaye JA. Overall survival, adverse events, and economic burden in patients with chronic lymphocytic leukemia receiving systemic therapy: real‐world evidence from the Medicare population. Cancer Med. 2021;10(8):2690-2702. doi:10.1002/CAM4.3855
3. Parviz M, Brieghel C, Werling M, et al. Post-treatment infection prediction in CLL using domain adaptation of lymphoma electronic health records. Acta Oncol. 2026:65:109-118. doi:10.2340/1651-226X.2026.44569
4. Levinson T, Wasserman A. C-reactive protein velocity (CRPv) as a new biomarker for the early detection of acute infection/inflammation. Int J Mol Sci. 2022;23(15):1-10. doi:10.3390/IJMS23158100