-Author name in bold denotes the presenting author
-Asterisk * with author name denotes a Non-ASH member
Clinically Relevant Abstract denotes an abstract that is clinically relevant.

PhD Trainee denotes that this is a recommended PHD Trainee Session.

Ticketed Session denotes that this is a ticketed session.

2233 Time Series Deep Neural Network Identifies Lymphoma Patients Suitable for CAR-T Cell Therapy Using EHR Data

Program: Oral and Poster Abstracts
Session: 803. Emerging Tools, Techniques, and Artificial Intelligence in Hematology: Poster I
Hematology Disease Topics & Pathways:
Artificial intelligence (AI), Lymphomas, Diseases, Lymphoid Malignancies, Technology and Procedures, Machine learning
Saturday, December 7, 2024, 5:30 PM-7:30 PM

Meihua Long, PhD candidate1,2*, Fan Jia1,2*, Jili Deng3*, Jianing Zhang2*, Ziran Zhao, PhD4,5, Yan Xie2*, Lan Mi2*, Yan Hou1,2* and Yuqin Song, MD2

1Department of Biostatistics, Peking University, Beijing, China
2Key Laboratory of Carcinogenesis and Translational Research (Ministry of Education), Department of Lymphoma, Peking University Cancer Hospital & Institute, Beijing, China
3Department of Oncology, Sichuan Cancer Hospital and Institute, Chengdu, China
4National Clinical Research Center for Chinese Medicine Acupuncture and Moxibustion, Tianjin, China
5First Teaching Hospital of Tianjin University of Traditional Chinese Medicine, Tianjin, China

Introduction

The efficacy of CAR-T cell therapy varies among lymphoma patients despite great advantages in sustaining remission and extending survival. Plenty of clinical markers have been identified as indicators for adverse events after CAR-T, but not as many have been revealed for forecasting the response to CAR-T therapy. Since lymphoma patients often went through many lines of therapy before receiving CAR-T treatment, clinical data from multiple time points might be indicative of CAR-T therapeutic effects. Here we proposed a times series deep neural network to predict the short-term remission response and long-term survival using all the patient’s medical record. The purpose of this study is to develop a predicting tool to assess whether lymphoma patients are likely to response to CAR-T therapy, thereby mitigating the financial cost for those unlikely to benefit from the treatment.

Methods

We retrospectively analyzed 96 patients with B-cell lymphoma who were treated with CD19-targeted CAR-T therapy at Peking University Cancer Hospital and Peking University International Hospital from January 2015 to June 2023. Demographic characteristics, hematological, hepatological, renal and immune markers that were examined from first line therapy to 3-month post CAR-T treatment were retrieved from medical record. The safety of CAR-T therapy were monitored for all patients, and the treatment response was assessed using PET-CT scans at before and after treatment. The Lugano Classification was applied to integrate PET-CT findings with clinical data to determine overall response as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD). The data were structured into before and after each line of therapy. The small scale Long Short-Term Memory (LSTM) model was trained to predict the 3-month response and long-term survival to CAR-T therapy, compared to using logistic regression model on pre-treatment data only.Permutation feature importance was applied to add interpretability of the model. To facilitate intuitive understanding and clinical usage of the model, we built a CAR-T prognostic score based on the LSTM model’s most important 5 features on pre-treatment data, and decision tree was used to determine cutoff values.

Results

Our LSTM model achieved better performance in both short term and long-term response prediction compared to logistics regression. For 3-month response prediction, our LSTM model achieved an average predicted probability of remission of 0.75 compared to 0.65 for the logistics regression model. For long term survival prediction, our LSTM model achieved an average predicted probability of success of 0.79 compared to 0.66 for the logistics regression model. Of the 100 biomarkers input in the model, uric acid, fibrinogen, creatine kinase, Lactate dehydrogenase (LDH), sodium, chlorine, mean corpuscular hemoglobin concentration (MCHC), hemoglobin, platelet, and alkaline phosphatase were discovered to be the important features predictive for the 3-month treatment response according to feature permutation method. Our CAR-T prognosis score on failure in treatment response increased by one point, when each of the following conditions was met before CAR-T treatment: fibrinogen <=121.8 mg/dL, LDH >281.5 U/L, MCHC <=327.5 g/L, platelet count < 45.5 g/dL and uric acid >590.5 µmol/L.

Conclusions

By incorporating time series data of previous lines treatment, our model improves the prediction on CAR-T cell therapy remission compared to only using the cross sectional data. Further work to validate our model on an external cohort or a prospective study is warranted.

Disclosures: No relevant conflicts of interest to declare.

*signifies non-member of ASH