Núria Buil-Bruna, Tarjinder Sahota, Helen Tomkinson
Quantitative Clinical Pharmacology, Early Clinical Development, IMED Biotech Unit, AstraZeneca, Cambridge, UK
Objectives:
Neutropenia is a common dose-limiting toxicity in oncology patients following either chemotherapy, radiotherapy or targeted therapies. Such toxicity risks severe complications and can cause dose reductions/interruptions or delay in starting a new cycle of treatment, potentially worsening treatment outcome. It has been suggested that the monocyte count either at baseline [1] or the rate of decline after start of treatment [2:4] could be a predictor of future neutropenia events. However, to the best of our knowledge, there has not been a systematic multivariate assessment of monocyte features as early predictors of neutropenia. The purpose of this work is to produce a classifier for safety monitoring to predict subjects at risk of developing neutropenia in their next visit.
Methods:
A total of 880 absolute neutrophil counts (ANC) and absolute monocyte counts (AMC) (obtained in average every 4 days) were available from 73 patients. Out of these 73, 15% (11 patients) experienced neutropenia at least grade 3 (ANC≤1 x 10^9 cells/L). Observations after first neutropenia event were excluded.
Three modelling approaches were compared to predict neutropenia: 1) univariate cut-offs 2) multivariate decision trees 3) other machine learning algorithms (support vector machines linear and radial, random forest and stochastic gradient boosting). The analysis was performed in R using the caret package. Neutropenia status at each visit was modelled using predictors derived from ANC and AMC from previous visits. Derived predictors were: baseline cell counts, cell counts from the three previous visits and slope between visits. To evaluate the predictive capability of AMC, models were assessed with and without AMC predictors. We assessed the predictive performance of each model via Monte Carlo cross-validation (n=100) using precision and recall on the hold-out datasets.
Due to the imbalanced nature of the data, the following approaches were used; 1) neutropenia events in train datasets were up-sampled to 50% prevalence; 2) the probability threshold for classifier cut-offs were investigated as hyperparameters using a separate bootstrap cross-validation (n=100) 3) the F2-score (a measure of precision and recall, weighing recall higher than precision) was used to optimise hyperparameters.
Results:
All models significantly improved cross-validation performance when AMC features were included. The three top models were multivariate decision trees (F2-score = 0.61, precision=0.39, recall=0.71, specificity = 0.8), support vector machine (F2-score = 0.58, precision=0.32, recall=0.73, specificity = 0.68), and stochastic gradient boosting (F2-score = 0.55, precision=0.54, recall=0.56, specificity = 0.90). The univariate cut-off model was outperformed by all algorithms. The multivariate decision tree was therefore selected. The most impactful feature in the decision tree was the change in AMC from baseline in the visit prior to observed neutropenia followed by the corresponding ANC. Interestingly, baseline AMC and ANC were found not to increase predictive performance of the decision tree classifier.
Conclusions:
We have demonstrated AMC significantly improve performance in predicting neutropenia. However, our results show that obtaining predictions with good operating characteristics for use in the clinic can be challenging, even when correlated variables have been identified. Although the more advanced machine learning algorithms provided the potential for improved performance, they come with additional practical and regulatory hurdles to implement due to the software requirement. The easy implementation of a decision tree may be a useful tool for clinicians to guide monitoring time of ANC in those patients with high risk of neutropenia. These findings require external validation with a larger dataset.
References:
[1] Moreau, Michel, et al. “A general chemotherapy myelotoxicity score to predict febrile neutropenia in hematological malignancies.” Annals of oncology 20.3 (2009): 513-519.
[2] Sato, Itaru, et al. “Prediction of docetaxel monotherapy-induced neutropenia based on the monocyte percentage.” Oncology letters 3.4 (2012): 860-864.
[3] Kondo, Mutsumi, et al. “Early monocytopenia after chemotherapy as a risk factor for neutropenia.” American journal of clinical oncology 22.1 (1999): 103-105.
[4] Ouyang, Wen, et al. “The change in peripheral blood monocyte count: A predictor to make the management of chemotherapy-induced neutropenia.” Journal of cancer research and therapeutics 14.10 (2018): 565.
Reference: PAGE 28 (2019) Abstr 8982 [www.page-meeting.org/?abstract=8982]
Poster: Drug/Disease Modelling - Oncology