Data management and exploration
In order to support clinical decision-making for patient-specific therapy planning, our prototype unifies data management, data description in the form of visualizations, and patient-specific predictions based on mathematical disease models. Figure 2 illustrates corresponding features and information flows of our prototype software.
At the Data management layer, our framework supports the management of patient identifying data as well as of pseudonymized payload data including medical and core data. If a physician has the permission to access patient identifying data (see access control) the software allows retrieval of de-pseudonymized medical data. In contrast, any other user, such as a documentarist, modeler, system administrator, etc. has per default no access to patient identifying data. However, depending on the defined access rights, users are allowed to retrieve particular pseudonymized medical data. Furthermore, our framework provides access to diagnostic procedures, diagnoses, treatment information (e.g. details on drug types and dosing) or other evaluation data in pseudonymized form. Depending on permissions, users are allowed to add, modify, (soft)Footnote 2 delete core data, medical data, and patient identifying data as explained in the section Access control.
To preserve data structure and to guarantee a high quality, we designed predefined data entry forms. However, to allow for flexibility, these can be customized by users on the basis of core data definitions. The core data sets define authorized entries of medical and (within pre-defined access rules) patient identifying data. As an example, when inserting leukocyte values from the peripheral blood for the first time, one has to define the core data “blood count” as screening type, “leukocyte” as diagnostic parameter, and the measurement unit, e.g. “109/l”. Thereafter, these entries are available in drop-down fields for insertion and editing. Additional information (i.e., metadata), such as date and time, screening status, order no., etc. are added automatically by the system or can be added in a user-defined way.
For the visual data description/exploration, we apply interactive Shiny routines. Visualization of monitoring parameters, e.g. time courses of leukocyte or platelet counts, or the proportion of Philadelphia-positive cells as well as BCR-ABL1 transcript levels, can (optionally) be presented together with therapy details and reference values/ranges. These visualizations help physicians to get a faster and more detailed overview of therapy conditions and corresponding patient responses (see Fig. 2: Data management / Data visualization).
Model-based decision support
On top of the data management and visualization features, our framework provides predictions, generated by mathematical models or computer simulations, to aid the physician’s decision making by complementing medical data with another level of information. The availability of model predictions might also help to communicate therapeutic decisions or potential alternative treatment scenarios to the patient and, therefore, to improve treatment compliance.
Based on the structured and visually presented clinical data, the physician can identify uncertainties or open questions that hamper a clear-cut therapeutic decision, such as expected patient-specific toxicity of treatment or necessity of individual treatment adaptations according to expected response. Such questions can then be addressed e.g. by simulating different treatment options for the same patient and generating corresponding model predictions “on-the-fly”. With the presented framework, the physician can specify the parameter settings (e.g. potential drug dose amendments, alternative application schedules or prediction intervals) directly within the GUI of the data management platform. Once the parameters have been selected, a simulation run can be started (Fig. 2: Simulation) and the patient-specific predictions are graphically presented together with the corresponding clinical data (Fig. 2: Model Visualization). The physician can appraise and potentially use this integrated information (data + model prediction) to arrive at his/her therapeutic decision.
It is also possible to generate model predictions for a spectrum of different parameter settings (e.g. a range of potential drug doses) to study potential effects sizes or sensitivities of expected patient-specific responses. All these “virtual treatments” (i.e. the MAGPIE project/job IDs of these particular simulations and the corresponding simulation data) are managed in the payload database (see Fig. 1) and are therefore always and completely reproducible.
Example applications / use cases
To demonstrate the functionality of our framework as a model-based clinical decision support system, we present two prototypic applications.
Prediction of treatment-response dynamics in CML (use case 1)
First, we implemented an example to illustrate decision support for CML patients under continuous tyrosine kinase inhibitor (TKI) therapy. In this disease, the level of the aberrant BCR-ABL1 mRNA in the peripheral blood provides a surrogate measure for tumour/leukaemia load, and is in routine clinical use for monitoring the patient’s treatment response ([8, 20, 21]). Our software visualizes the BCR-ABL1 level over time, and allows annotation of the raw data points by different reference values and/or actual treatment modalities, such as TKI type/dose (Fig. 3a). Targeting the individual data point with the mouse cursor will provide additional information about this particular measurement, e.g. quality criteria such as the underlying copy number or whether this data points has been standardized on the international scale (IS).
Applying our established mathematical CML model ([5, 8]) to an individual patient time course (i.e., estimating the model parameters from BCR-ABL1 measurements) allows to derive and visualise patient specific predictions about the expected future treatment response (Fig. 3b). Especially, the estimated abundance of residual leukemic stem cells is not accessible in the clinic and provides additional model-derived information, e.g. in the case that treatment cessation is considered. Currently, the CML model provided assumes a fixed TKI dose. For mathematical detailed about the modelling and the parameter estimation, we refer to [8].
Additionally to the BCR-ABL1 levels together with the model predictions, the user is able to access further clinical parameters that are available for this patient. These can be retrieved and visualized via the graphical menu (c.f. Fig. 3a).
Prediction of thrombocytopenia under cytotoxic chemotherapy (use case 2)
As a second illustrative example, we implemented the individualized mathematical model of human thrombopoiesis, applied to patients with aggressive NHL treated with six cycles of a combination therapy of four to five cytotoxic drugs, i.e. applying the CHOP / CHOEP chemotherapy regiments studied in [22]. In these protocols, cycle duration is either 14 or 21 days. Patients treated with these chemotherapies are at high risk for developing life-threatening haematotoxicity during the course of the therapy [23]. Predicting which of the patients suffer these severe conditions is of high clinical relevance to take countermeasures such as prophylactic hospital stay, postponement of therapy or reduction of chemotherapy dosage.
The mathematical thrombopoiesis model ([17]) uses individual platelet time course data and the applied therapy schedules together with population data from the literature to estimate individual model parameters. These parameters can be used to predict individual future platelet counts of this particular patient. This includes simulations of treatment adaptations, aiming to minimize thrombocytopenia while maintaining sufficient treatment efficacy.
The simulations can be configured by selecting treatment options such as dosing of drugs. Results are visualized within the GUI of our framework (Fig. 4). In brief, available clinical data of a specific patient, including basic patient characteristics, planned treatment protocol and platelet counts during therapy can be directly assessed and visualized (Fig. 4a). Boundaries of the different degrees of thrombocytopenia are also provided, if desired by the user. So far, available platelet counts and prior data are used to derive individual parameter estimates. Based on this parameter set, the user can perform simulations of future treatments including adaptations of the originally planned protocol. For this purpose, the start of the next therapy cycle can be shifted by a specified number of days. Moreover, doses of all cytotoxic drugs can be adapted, or the software determines a dose factor to tune the degree of thrombocytopenia to a tolerable level. Corresponding model predictions can be displayed for a specified follow-up time together with the currently available data (Fig. 4b). This model-based assessment of different treatment options supports clinical decision-making regarding timing and dosing of the next therapy cycle.