1. What is the current knowledge on the topic?

Lapatinib was approved in China to treat patients with HER2(+) metastatic breast cancer in combination with capecitabine based on a single-arm, open-label study (EGF10949). Two dose regimens are commonly recommended for lapatinib, 1,250 mg of lapatinib in combination with capecitabine and 1,000 mg of lapatinib in combination with trastuzumab. Under the recommended dose regimen, lapatinib can be well tolerated with minimal avoidance of drug toxicities.

2. What question did this study address?

In this study, we established a deep learning model to predict the lapatinib dose based on important influencing variables from real-world evidence, resulting in getting the optimal individualized dose regimen.

3. What does this study add to our knowledge?

This study provides a new perspective and guidance for lapatinib dose administration where few studies focused on individualized lapatinib dose treatment in breast cancer patients previously.

4. How might this change clinical pharmacology or translational science?

Models based on machine learning and deep learning methods could help clinicians treat breast cancer patients with individualized lapatinib dose regimens to get the optimal effect and reduce adverse events.

Lapatinib is a selective inhibitor of the tyrosine kinase receptor and human epidermal growth factor receptor-2 (HER2) and has activity in HER2-overexpressing breast cancer (1–3). By binding to the ATP-binding site of the receptor’s intracellular domain, lapatinib blocks HER2 tyrosine kinase activity leading to inhibition of tumor cell growth (4). After the progress with anthracycline, taxane, and trastuzumab in China, lapatinib has been introduced for the treatment of advanced/metastatic HER2(+) breast cancer (1, 4–6). Two dose regimens are commonly recommended for lapatinib, which can be optimally tolerated (7). For patients with advanced HER2(+) breast cancer progressing with therapy with anthracyclines, taxanes, and trastuzumab, it is recommended to administer 1,250 mg of lapatinib in combination with capecitabine (4). For patients with metastatic HER2(+), hormone receptor(-) breast cancer upon progressing with therapy with trastuzumab and chemotherapy, it is recommended to administer 1,000 mg of lapatinib in combination with trastuzumab (4). Under the recommended dose regimen, lapatinib can be well tolerated with minimal avoidance of drug toxicities, which are skin rash and diarrhea predominantly (6, 8–10). Therefore, a promising model to predict an appropriate individualized dose regimen is important to get a balance of lapatinib efficacy and toxicities to improve the treatment outcome.

With the rapid development of information technology, real-world study has become an important data source for clinical research (11). Most real-world studies use information from electronic medical records, examination data, and follow-up records during diagnosis and treatment. Real-world study is a process of data mining, model building, and clinical feature data extraction. The main advantages of real-world studies include rich evidence resources, good external validity, individualized program application, and being closer to clinical practice (12, 13). Compared with conventional modeling methods, machine learning and deep learning techniques have indubitable advantages in dealing with real-world evidence, such as the following: (1) machine learning and deep learning can deal with more complex, high-dimensional, and interactive variables, which is lacking in traditional models, and (2) machine learning and deep learning models have stronger generalization and better accuracy than traditional models (14–16). Recently, some algorithms with more sophisticated principles have been developed, such as eXtreme Gradient Boosting (XGBoost), light gradient boosting machine (LightGBM), Categorical Boosting (CatBoost), Gradient Boosting Decision Tree (GBDT), and TabNet, which have been highly recognized in algorithm competitions (17–21). Recently, the application of machine learning and deep learning techniques based on real-world study has been a trend, such as a novel prognostic scoring system of intrahepatic cholangiocarcinoma with ensemble machine learning algorithms (XGBoost, random forest, and GBDT), a prediction model of tacrolimus blood concentration in patients with autoimmune diseases using XGBoost, a novel vancomycin dose prediction model through XGBoost, and warfarin maintenance dose prediction through LightGBM (22–25). Many studies have demonstrated the advantages of machine learning algorithms over traditional statistical methods. With the increasing number of input subject data, machine learning and deep learning models can continually optimize parameters to achieve better performance and practicality.

In order to achieve a balance of drug efficacy and toxicities, an appropriate dose regimen is important for patients’ treatment outcome. In this study, we aim to establish a model based on machine learning and deep learning techniques to predict the lapatinib dose based on important influencing variables from real-world evidence, resulting in getting the optimal individualized dose regimen.

Lapatinib was approved in China to treat patients with HER2(+) metastatic breast cancer in combination with capecitabine based on a single-arm, open-label study (EGF10949), which evaluates the drug efficacy and safety (1, 27). Our study focused on the establishment of a prediction model for lapatinib dose, mainly evaluating two dose regimens (1,000 and 1,250 mg). We used TabNet, a leading-edge deep learning technique, to construct the prediction model with good performance (accuracy = 0.82 and AUC = 0.83). Afterward, important variables that strongly correlated with lapatinib dose were ranked via an importance score, including treatment protocols, weight, number of chemotherapy treatments, and number of metastases. Lastly, the confusion matrix was used to validate the model; it can be seen that the dose regimen of 1,000 mg lapatinib had a precision of 88% and a recall rate of 64%, and the dose regimen of 1,250 mg lapatinib had a precision of 82% and a recall rate of 95%.

As a novel deep learning technique, TabNet uses a sequential attention mechanism to choose a subset of meaningful features to process at each decision step, enabling interpretability and more efficient learning as the learning capacity used for the most salient features (21). Additionally, based on retaining the end-to-end and representation learning characteristics of deep neural networks, TabNet also has the advantages of tree model interpretability and sparse feature selection (28). Other studies based on real-world evidence show that TabNet outperforms ensemble tree-based algorithms, since it can process a highly nonlinear relationship with its depth, without overfitting due to instance-wise feature selection (21). In this study, we applied TabNet, super TML, Wide&Deep, and ANN, which are algorithms with good predictive ability in various network algorithm types. After comparing the different network models, the TabNet model shows the best prediction performance (Table S1). Later, we will increase multicenter data. With larger data volume, we will build different network models and explore the optimum one to predict the dose of lapatinib to assist clinical medication.

As one of the most important variables influencing lapatinib dose, protocols of combination therapy accounted for a different proportion in this study as follows: protocol_1, protocol_2, protocol_3, and protocol_4 accounted for 67.11%, 16.11%, 6.71%, and 10.07%, respectively. The combination of lapatinib + capecitabine (protocol_1) was one of the common recommended regimens for patients with HER2(+) breast cancer with prior treatment of taxanes, anthracyclines, and/or trastuzumab (29, 30). Clinicians generally used lapatinib at a dose of 1,250 mg daily continuously plus capecitabine at a dose of 2,000 mg daily, and the combination can achieve superior treatment efficacy than capecitabine monotherapy (29, 31). In addition, a study on Japanese breast cancer patients found a drug–drug interaction and pharmacokinetics interaction between lapatinib and paclitaxel, as the AUC and Cmax of these patients given the combination therapy were affected (32). In Rezai et al.’s study, a pharmacokinetic interaction was found between lapatinib and vinorelbine, and the use of lapatinib can remarkably decrease the vinorelbine clearance (33). It has been proven that therapy of lapatinib combined with other drugs can commonly improve the time to progression and/or achieve longer survival time in these patients (29, 34, 35). Normally, previous studies focused on the efficacy and safety of combination therapy; few investigated the effect on dose regimen. Further research is needed to confirm the relationships between combination therapy and drug doses.

For the physiological features, previous studies found that patients with lower weight were more likely to have higher lapatinib plasma levels (P = 0.055) (36). The researchers believed that patients without fasting and lapatinib dose were not adjusted to low body weight during drug intake, which would lead to increased lapatinib body levels (36). The number of chemotherapy treatments was identified as an important influencing variable for lapatinib dose, and a previous study showed that three or more prior treatments strongly correlated with worse survival (1). Additionally, some studies found that organ metastases (such as liver and brain metastases) show a strong relationship with morbidity, mortality, and survival rate in breast cancer patients, which may affect the lapatinib dose regimen (37–39). In this study, the number of metastases has a remarkable effect on lapatinib dose whose correlation was previously investigated by few studies, which warranted further research.

According to the confusion matrix results, the classifier correctly identified 82% of patients using the 1,250-mg lapatinib regimen and 88% of patients using the 1,000-mg lapatinib regimen, indicating a remarkable prediction performance. Nevertheless, the sample size in the training and test cohorts was small. Large samples are required to verify this result.

Our study endeavored to build a dose prediction model via machine learning and deep learning methods, which could mine deep data based on real-world evidence. Through a comparison of different algorithms, TabNet was selected to establish the model based on the strength of its predictive ability. To our knowledge, few studies focused on lapatinib dose prediction previously, and this study provides a new perspective and guidance for lapatinib dose administration with a more concise and accurate model. Compared with conventional models, machine learning and deep learning models mine and use unexploited variables to cover the shortage of clinical experience from the real world. One limitation in this study was the limited sample size that affected the model to further optimize the performance. In future, more real-world evidence should be added in the model to optimize its performance, and larger prospective clinical studies will be needed to investigate the further interactions between different variables and lapatinib dose.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Ethics Committee at Fudan University Shanghai Cancer Center (FUSCC) (No. 2016-106-1159-K1). The patients/participants provided their written informed consent to participate in this study.

ZY and XY led the research and provided medical support. FK analyzed and interpreted the data. HYL, JZ, and ZW analyzed the data and provided technical support. HL and XH wrote the manuscript. HW and FG provided methodological guidance. QZ provided technical and medical guidance. All authors contributed to the article and approved the submitted version.

FG was supported by the National Key Research and Development Program (2020YFC2005502, 2020YFC2005503). QZ was supported by the Clinical Research Plan of SHDC [SHDC2020CR3085B].

Author JZ and FG are employed by Beijing Medicinovo Technology Co. Ltd., China. Author XH is employed by Dalian Medicinovo Technology Co., Ltd., China.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.