Machine Learning for Prediction of Serious Infections in Patients with Treatment-Naïve Lymphoma; A Population-Based Study of 701 Patients from the North Denmark Region

Purpose: First-line treatments with immuno-chemotherapy for lymphoma are associated with risk of life-threatening infections due to treatment-related immunosuppression, but only a minority develops severe infections that require hospital admission. In most cases, patients recover quickly without complications. Accurate identification of patients at risk of serious infections would be of substantial value for patients and society, for example through reduction in number of hospital admissions. Current risk stratification methods are based on simple clinical features such as fever, blood pressure, and symptoms, but were not developed specifically for use in lymphoma. The primary objective of this study was to develop a machine learning (ML) model that can predict risk of serious infections during or after first-line treatment for lymphoma.

Methods:

This study utilized population-based data from the North Denmark Region. Adult patients diagnosed with lymphoma between 2013-2023 and registered in the Danish Lymphoma Registry (LYFO) were included if treated with CHOP (or CHOP-like), ABVD, BEACOPP, CVP, BENDAMUSTIN chemotherapy backbones in first line. Furthermore, structured data were retrieved from patient records and registries, including vital measurements (pulse, temperature, blood pressure, height, weight, GCS, and saturation), clinicopathologic features obtained at diagnosis, treatment information, hospital contact patterns, and laboratory data. An infection related hospitalization was defined as a hospitalization shorter than 30 days with an infection related ICD-10 discharge code during or within 12 months after first-line treatment. Patients were censored at relapse/death.

An XGBoost model was developed using stratified K-Fold Cross-validation stratified for high-risk/low-risk patients in all folds. A feature matrix was constructed using multiple time-related lookback windows to capture information in the data at various times in the patient’s history. The model was blinded for all data after each prediction point. Patients were censored at relapse or 12 months after completing first-line treatment. A high-risk infection was defined as a hospital admission with presence of one or more of the following events during the first 48 hours: temperature > 39˚C or < 36˚C, systolic blood pressure < 100 mmHg, saturation < 90%, respiratory frequency > 22 breaths/minute, pulse > 120 beats/minute, or GCS < 15. Positive blood culture or death at any time during hospitalization were also considered high-risk.

Results:

Out of 1,408 patients screened, 701 patients met the inclusion criteria. The median age at diagnosis was 67 years (IQR: 55-75), with 59% being male. The most used treatments were CHOP (or CHOP-like) (64.7%), ABVD (13%), CVP (9.7%). A total of 1,084 hospitalizations related to infections were identified, including 382 high-risk infections according to the listed criteria. The feature selection process generated 8,315 features based on the different datasets using varying relevant lookback windows and aggregators (such as minimum value and mean value in each lookback window). At a probability threshold of 0.4 the mean performance of five folds was: ROC-AUC: 0.89, F1: 0.75, Sensitivity: 0.86, Specificity: 0.75, NPV: 0.90 and PPV: 0.66. Analysis of the Shapley Additive (SHAP) values showed that the value last diastolic blood pressure measurement and mean pulse in the past 90 days, were the two most influential features. Aiming for an NPV of >95%, the model was able to correctly identify 363 low-risk admissions using prediction cut-off of 0.18. With a mean hospital stay of 5.5 days in the study population, this would have amounted to a total reduction of 1997 bed days if low-risk patients would have been treated in an outpatient setting. With this threshold there was a total of 21 false negatives predictions, i.e., patients predicted as low-risk but with one or more high-risk events and 294 were false positives, i.e. predicted high-risk but with no subsequent high-risk defining events.

Conclusion: We found our ML-based model showed great results for predicting high-risk patients, predicting 361 correct high-risk and 363 correct low-risk, hence saving the bed occupancy resources for the low-risk patients, and reducing unnecessary hospitalizations. This model is still in preliminary phase and need external and prospective validation before being ready for implementation.