Optimizing the use of blood based biomarkers to improve the clinical management of advanced non-small cell lung cancer

In recent decades progress has been made in the treatment of non-small cell lung cancer (NSCLC). The introduction of targeted therapies resulted in a survival benefit compared to traditional chemotherapy in patients presenting with actionable genomic mutations. In patients without actionable genomic mutations, immunotherapy has become the preferred treatment. Immunotherapy activates the immune system. However, this process can result in inflammation in various organs. This process is referred to as an immune-related adverse event (irAE). Despite providing a survival benefit in patients responding to immunotherapy compared to chemotherapy, a substantial group does not respond to immunotherapy. Due to the high costs of immunotherapy and the probability of developing severe irAEs, it is necessary to identify patients who will not show clinical benefit in an early stage of therapy. Also, early identification of non-response allows for the initiation of a potentially more beneficial treatment earlier, i.e., before further disease progression, thereby improving the likelihood of patients being eligible for entering clinical trials. In current clinical practice, blood samples are used to obtain information on both the development of irAEs and therapy response. The use of body fluids for laboratory testing is referred to as a liquid biopsy. The advantage of liquid biopsies is the ease of access to the biological material, low patient burden and costs. In current practice, liquid biopsies are used to obtain serum tumor marker (STM) measurements, which are known to reflect tumor mass and thus provide information on treatment response. Other laboratory tests can provide information on organ function and can therefore be employed to detect and grade irAEs. Although these biomarkers have already been discovered in recent decades, guidance regarding the use of these biomarkers in immunotherapy treatment of NSCLC patients is lacking. Both monitoring of therapy response and the detection of irAEs poses their own questions and challenges. Although the incidence of irAEs is low, irAEs are known to severely impact treatment schedules and outcomes. Therefore it is necessary to identify irAEs in an early stage and adjust the treatment accordingly. In the Netherlands Cancer Institute, a wide laboratory test panel is performed before each new round of immunotherapy to early detect irAEs. However, the specificity of the individual tests is low since test results can be influenced by other medication, inflammation, or a non-cancer-related disease. The high test frequency and large size of the test panel could result in a large stream of abnormal test results not related to an irAE. Although the test panel is used in current practice, some questions remain. What is the added value of the individual test in the test panel? How to optimize a test panel used for the diagnosis of low-incidence events?
In current practice immunotherapy response evaluation strongly relies on radiological imaging using computed tomography (CT) scans and Response Evaluation Criteria in Solid
Tumors (RECIST). These response evaluation criteria are predominantly based on evaluating changes in tumor size of selected lesions. Therefore progression outside of the scan area or changes in non-selected lesions can be missed. STMs are known to reflect total tumor mass, and changes in STM measurements might better reflect therapy response. Despite the potential value of STMs in response monitoring, consensus is lacking. However, some physicians use STM measurements in daily practice. While the use of STMs in the assessment of immunotherapy response might hold potential, several questions have been raised. Could STMs be used to predict response in NSCLC patients treated with immunotherapy? How can sequential STM measurements be used in a prediction model? What is the potential clinical impact of introducing a prediction model in clinical practice? Aiming to make a substantial contribution to answering these questions, the first part of this thesis aims to optimize the detection of irAEs using a simulation modelling approach, while the second part of this thesis is directed towards the development and evaluation of a response prediction model. The first part of this thesis (Chapters 3 and 4) aims to optimize the detection of irAEs by using a simulation modelling approach. To reflect the clinical pathway in a simulation model, it is necessary to create a simulation model that captures both disease progression and the development and detection of irAEs. These are two independent, asynchronous, parallel processes. Chapter 3 describes how a simulation model based on the principle of timed automata (TA) can be used to simulate the clinical pathway of immunotherapy treatment in NSCLC patients. A TA-based model is especially suited for this purpose since it is composed of multiple templates, each reflecting a specific sub-routine. Synchronization channels allow communication between templates enabling the simulation of asynchronous parallel processes. Results from Chapter 3 show that the TA-based model is able to provide results similar to actual patient data after calibration. Moreover, results indicate that the specificity of the laboratory tests is the primary driver influencing immunotherapy continuation. However, the model as presented in Chapter 3, is not designed to allow for the assessment of all included laboratory tests independently since all laboratory tests were grouped per irAE, and the interpretation of test results by a physician was incorporated in the overall test accuracy. To allow for the assessment of individual laboratory tests, the model was further developed in Chapter 4. In the updated model, test results and interpretation of these results by a physician are divided into two separate processes. Results of Chapter 4 show that when multiple tests are employed to detect the same irAE, reducing the test panel size might be possible without introducing a delay in the time to irAE detection. Other analyses indicate, however, that a reduction in test frequency likely delays irAE detection and potentially results in a higher grade irAE at diagnosis. The second part of this thesis (Chapters 5, 6, 7, and 8) describes the development of prediction models aimed to predict immunotherapy non-response and the evaluation of the most promising prediction models in both an external validation cohort and a simulation study. Since STMs are known to reflect tumor volume, it was hypothesized that changes in STM measurements during early treatment might be predictive of therapy response. Early identification of non-response enables initiation of an earlier treatment switch, allowing to start a potentially more beneficial treatment earlier. In current practice, the first moment of response evaluation is after six weeks. Therefore, sequential STM measurements taken during the first six weeks of immunotherapy were selected as input for the prediction model. A high specificity (>95%) was deemed necessary since the outcome of the prediction aims to inform a therapy switch. Chapter 5 illustrates how prediction methods with varying levels of complexity can be used to predict therapy response using sequential STM measurements. Results of this study indicate that CYFRA proves the most predictive information amongst the STMs evaluated. The highest accuracy on the internal validation was achieved by using a recurrent neural network (RNN) applied to sequential CYFRA measurements, resulting
in a sensitivity and specificity of 38.6% and 100%, respectively. Since all STMs evaluated in Chapter 5 have shown potential predictive power, it was hypothesized that combining multiple STMs in a single prediction model might improve the predictive power. Therefore, multiple prediction models combining different combinations of sequentially measured STMs were evaluated in Chapter 6. Results from this study show that a boosting model combining CYFRA and CEA as input achieved the highest sensitivity while maintaining a >95% specificity on internal model validation. The sensitivity and specificity achieved by the boosting model were 59.4% and 96.6%, respectively.
The five most promising prediction models from Chapter 6 were selected for external validation. This external validation is described in Chapter 7. To allow for external validation, data was collected from a new cohort of NSCLC patients treated with immunotherapy at the Netherlands Cancer Institute or the Radboud University Medical Centre. A Random Forest (RF) model using CYFRA, CEA, and NSE as input achieved the highest sensitivity (38.1%) while maintaining a >95% specificity (95.9%). By combining the prediction model with results of the CT-scan performed in week 6, the positive predictive value increased to 100%. However, this resulted in a drop in sensitivity to 20.6%. Despite limited sensitivity the results of this study show a clinically relevant difference in progression-free survival (PFS) between patients with different response predictions (i.e. response vs. non-response). A median PFS of 1.8 months was found in patients in whom the combined prediction model predicted non-response. In the patient group for which a response was predicted, a median PFS of 8.0 months was found. The external validation of the prediction model provides a first insight into the potential
clinical impact of introducing the prediction model into clinical practice. However, this study only provides limited information on the impact on treatment duration. Therefore,
in Chapter 8, a simulation study was performed to compare care as usual to a scenario in which the prediction model was implemented in clinical practice. The simulation study was designed to simulate the immunotherapy treatment and evaluate treatment duration and costs. Results indicate that, introducing the prediction model in clinical practice could reduce average immunotherapy treatment costs by €5.198 per patient. This reduction in costs is attributed to a reduction in the average immunotherapy treatment duration of 16 days per patient. The reduction in treatment duration consists of a reduction in overtreatment in 23.7% of patients and unintended undertreatment in 2.5% of patients. However, results from this simulation study also show the potential impact of unintended undertreatment. On average, in patients in whom unintended undertreatment occurred, a treatment switch was initiated 259.3 days early. Indicating the value of performing simulation studies to evaluate design criteria during early development. This thesis provides insight into how information acquired through liquid biopsies can be used to improve immunotherapy in NSCLC patients. Modelling experts can play a key-role in this development by developing flexible simulation models capable of dealing with parallel asynchronous processes, which results can be used to inform further clinical studies. Although STMs are currently used by some physicians in clinical practice, further adoption of these biomarkers might be improved by better guidance. Results from this thesis show how sequentially measured STMs can be employed to predict immunotherapy treatment response. Moreover, this thesis also shows how simulation models can be used to evaluate prediction models and guide further development. Overall, this thesis showcases how prediction models, in combination with flexible simulation models and multidisciplinary collaboration, can result in a more optimal use of already existing data sources.