AI predicts how kids with ITP will respond to rituximab infusion
Model predicts higher response in older children, girls
Artificial intelligence (AI) can accurately predict a response to the second-line therapy rituximab in children and adolescents with immune thrombocytopenia (ITP), according to a study.
“Our [machine learning] model can optimize the management of paediatric ITP undergoing rituximab therapy and assist clinicians in making informed decisions in resource-limited settings,” the researchers wrote.
The study, “Machine learning to forecast rituximab responses for paediatric immune thrombocytopenia: Forging a path towards personalized medical care,” was published in the British Journal of Haematology.
ITP is caused when self-reactive antibodies drive the destruction of platelets, the cell fragments that help the blood clot after an injury. Such autoimmune attacks lower platelet counts and increase the risk of bruising and bleeding.
First-line ITP treatment typically involves immunosuppressing corticosteroids and intravenous immunoglobulin therapy (IVIG). For patients who do not respond to these treatments, second-line options are available, including rituximab, which is sold under the brand name Rituxan and others, with biosimilars also available. Rituximab, given intravenously (into the vein), is designed to lower the levels of B-cells, the immune cells that produce antibodies, including the self-reactive ones that drive platelet destruction in ITP. Although the therapy is not formally approved for ITP, it’s commonly used off-label for treating autoimmune diseases.
Antibodies, bleeding severity among predictive factors
A research team in China used machine learning — a type of artificial intelligence that uses algorithms to learn from data and identify patterns — to predict the initial response to rituximab in children and adolescents with ITP.
The team extracted data from the medical records of 156 pediatric ITP patients, ages 1-16, who were treated with a low-dose rituximab regimen at Beijing Children’s Hospital from August 2020 to August 2023. Half of the participants responded to rituximab, meaning their platelet counts reached at least 30 billion platelets per liter within three months of treatment initiation.
Data including demographic and clinical information, as well as blood test results, were input into four machine learning algorithms. Each model was then evaluated for accuracy (overall correctness of predictions), precision (the accuracy of positive predictions), and recall (ability to identify true positives).
A machine learning model called multilayer perceptron (MLP) had the highest area-under-the-curve (AUC) — 0.88 — in its ability to accurately distinguish rituximab responders from non-responders. AUC is a number between 0 and 1 that indicates a test’s ability to differentiate between two states, with values closer to 1 reflecting better accuracy. The AUC for MLP’s precision was 0.82, and for recall it was 0.84.
Factors that predicted a positive rituximab response included levels of antinuclear and anti-thyroglobulin self-reactive antibodies, pre-rituximab corticosteroid response, and bleeding severity. Negative rituximab response factors included levels of anti-thyroid peroxidase antibodies, certain immune T-cells, and the duration of disease before rituximab treatment.
In a subgroup analysis that compared MLP-predicted rates with observed rates, the model predicted a higher response rate among patients older than 5 than in younger patients (55.1% vs. 47.4%). This finding aligned with the observed rates (55.1% vs. 44.8%), “effectively capturing age-related variability in rituximab response,” the team wrote.
The model predicted a higher response rate in girls than in boys, also matching observations, with a higher prediction accuracy for girls than boys (88.1% vs. 77.8%).
Among patients with mild bleeding, the predicted rituximab response rate was slightly higher than the observed rate, while for those with severe bleeding, the predicted rate was marginally lower. Despite these discrepancies, MLP maintained high prediction accuracies for rituximab response in patients with mild and severe bleeding (80.6% and 87.9%).
The predicted rate closely matched the observed rate for patients with newly diagnosed ITP lasting less than three months, and for those with persistent ITP lasting three to 12 months. For chronic ITP, or disease lasting more than 12 months, the predicted response was slightly higher than the observed rate, but it achieved the highest prediction accuracy (93.6%).
Overall, the MLP model exhibited strong predictive performance across all subgroups, with prediction accuracies ranging from 76% to 93.6%.
“The proposed [machine learning] model captures the complexity of ITP [disease mechanisms] and rituximab responses,” the researchers wrote. “Future work will focus on broader validation and integration of this model into clinical decision support systems to enhance personalized treatment strategies and improve outcomes.”
About the Author
