Jul 24, 2024
Expert interview with Dr Jan Roediger: From theory to practice
StimFit algorithm may increase efficiency of DBS
A new imaging-based algorithm suggests individual DBS settings and could simplify the clinical practice of deep brain stimulation in Parkinson’s disease
Neuroimaging has made rapid advances, allowing to precisely reconstruct electrode placement in deep brain stimulation (DBS). At the same time, advancements in DBS devices have widened the range of possible stimulation settings. The correct selection of stimulation parameters based on trial-and-error has become a time-consuming and challenging process. A recently published algorithm (StimFit) guiding DBS programming based on neuroimaging data could support clinical practice in the future of DBS.
We asked Jan Roediger, lead author of the study conducted within the collaborative research center SFB TRR 295 ReTune, to provide insight into his research and what next steps and challenges are needed for the promising results to reach clinical practice.
Dr Roediger, DBS is an established treatment option for patients suffering from Parkinson’s disease. What are the current challenges in the process of setting stimulation parameters?
Currently, programming DBS devices for Parkinson’s disease (PD) patients is a time-consuming, process that requires multiple in- and outpatient visits, over months or years. Still, the overwhelming number of possible stimulation settings modern DBS devices have to offer cannot be sufficiently explored by current manual procedures due to several challenges. These include differential delays in symptom response after stimulation onset, symptom fluctuations, medication adjustments and patient fatigue.
To account for these factors highly trained medical personnel and specialized centers are required. With the expected worldwide increase in PD prevalence in the coming decades, and the economic pressures and staff shortages facing healthcare systems, it is crucial to reduce the complexity of postoperative care.
What are the potential solutions to these problems? Are there already established methods available to tackle these challenges?
There are multiple promising scientific approaches for automated or semi-automated DBS programming, including the use of kinematic feedback or electrophysiological signals. Our approach focuses on the anatomical location of patients’ DBS-electrodes, using neuroimaging data, which is routinely acquired before and after surgery. This makes it particularly feasible to implement in clinical practice, as no additional data acquisition or hardware is required.
The idea of guiding DBS programming through neuroimaging-derived metrics was first proposed almost two decades ago when a concept called “Volume of Tissue Activated” (VTA) was introduced to estimate the anatomical regions affected by the stimulation. As a result, stimulation settings can be probed within the software (in silico), aiming at focusing stimulation on specific target structures while avoiding neighboring regions. This concept is already implemented in commercially available products, and several prospective studies have emphasized the advantages in terms of reducing programming time.
So commercially available products have already proven to be useful for programming of DBS devices. However, you and your group at Charité have developed a new software called StimFit. How does StimFit work and what are the key differences to the already established approaches?
Current approaches to image-guided DBS programming, although a significant improvement over earlier “blind” methods, have limitations. They rely on biophysical models, which attempt to mathematically describe cellular-level responses to DBS. These “bottom-up” models, originally developed to understand DBS mechanisms, were not designed to predict stimulation outcome. Their reliance on biophysical assumptions limits their flexibility and reliability in predicting optimal stimulation settings.
Given the dramatic developments in AI technology, data-driven approaches hold overwhelming potential for the future of medical technology. In DBS neuroimaging, we need to shift from biophysical assumptions to data-driven approaches to capture the relationship between electrode position, stimulation parameters, and clinical outcomes. This transition allows for outcome quantification and enables increasingly detailed models as more data becomes available, facilitating highly individualized treatments.
The StimFit algorithm represents a step in this direction. StimFit recommends optimal stimulation parameters based on each patient’s electrode location and disease characteristics. Using the Lead-DBS software, individual DBS electrodes are mapped onto a standard brain. StimFit then simulates thousands of different stimulation settings, predicting the symptom response and the likelihood of side effects of each simulated setting. This approach effectively transfers the task of testing stimulation settings from the patient’s bedside to a computer. With the design of the model, we aimed to address key AI concepts such as the bias-variance tradeoff and to rely as little as possible on unknown biophysical assumptions. We tested the algorithm on retrospective datasets to ensure that it could indeed predict stimulation outcomes in patients.
After developing, training and retrospectively validating the algorithm you and your research team at Movement Disorders and Neuromodulation, led by Professor Kühn aimed to find out whether the settings suggested by StimFit can compete with those determined through clinical testing. Can you give us a summary of the study’s findings?
We conducted a double-blind, randomized, crossover trial to compare the effectiveness of the StimFit algorithm with standard clinical programming (SoC) in determining optimal DBS settings for patients with Parkinson’s disease. We enrolled 35 patients who had already undergone DBS programming based on our center’s standard of care. These patients were randomly assigned to receive either StimFit-programmed DBS first and SoC-programmed DBS second, or vice versa.
We evaluated the effects of each setting and found that both stimulation settings improved motor symptoms equally well, with no clinically significant difference. Importantly, the SoC program was the result of multiple in- and outpatient visits during which stimulation parameters were adjusted according to our centers standard clinical procedures. StimFit on the other hand needed two hours of computational time to suggest settings based on electrode locations in a fully automated manner. The successful outcome of the study indicates that data-driven algorithms based on neuroimaging derived metrics could be used to suggest effective stimulation settings and significantly reduce programming time.
What additional steps are necessary for the promising results to reach clinical practice? What are the technical and legal challenges that need to be overcome?
StimFit is a scientific software designed to demonstrate the proof-of-concept of data-driven DBS optimization. Our initial goal was not to develop a commercial product but to contribute technical and conceptual innovations to the research field. This is also why we decided to release the code as open source. However, we are now considering steps for its successful clinical application due to the potential economic and therapeutic benefits.
The challenges we face include technical development, infrastructural compatibility, clinical integration, and commercialization.
First, technical development involves further improving the model. It is important though to balance the additional work on the model with the expected benefits, as ideas for improvement are inexhaustible.
Second, we need to integrate the algorithm within a hospital’s IT system and automate the processing pipeline. In my opinion this represents one of the biggest challenges in the whole process, as hospital IT infrastructures typically evolve organically over the years in response to new demands and technological advancements without a cohesive strategy. This creates a chaotic environment lacking standardization, modularity, and interoperability, making it difficult to develop a one-size-fits-all solution.
In parallel, additional clinical trials are necessary and questions need to be answered, such as how to adjust medication alongside the stimulation, how many different software-based settings should be tried in a patient, and how to adjust the settings if the outcome is unsatisfactory. We also need to investigate the long-term effects of these stimulation settings on the patient’s overall quality of life.
Finally, we must consider the steps for licensing, commercialization, and distribution of the software. StimFit must obtain approval as a medical procuct from regulatory bodies such as the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA). Establishing partnerships with healthcare providers, medical device companies, and technology firms will be necessary to facilitate the integration and distribution of StimFit. Continuous monitoring post-deployment is necessary to ensure ongoing safety and effectiveness, involving the collection of real-world data and making necessary updates based on user feedback and new research findings.
Only if all these areas are addressed comprehensively, StimFit can transition from a promising research tool to a widely adopted clinical solution to improve postoperative care in patients treated with DBS.
How do you imagine this technology to work in the future, lets say in 20 years from now? What additional technical innovations might play a role?
Integrating multiple biomarkers—such as genetic, metabolic, and even microbiome information—alongside digital biomarkers from kinematic, electrophysiological, and imaging data, is expected to improve characterization and differentiation of disease states and treatment of movement disorders, where current diagnostic criteria heavily rely on the clinical presentation. Given that we are interfering in the processes of the most complex biological structure on earth, a phenotypical classification of diseases seems like an almost absurd oversimplification. Hence, therapy will inevitably benefit from being tailored to individual neuroanatomical and neurophysiological characteristics, potential comorbidities, and the patient’s subjective needs. Other technical innovations, such as wearable technology and remote monitoring of real-time data on patients’ conditions will contribute to these developments.
The collection of high-quality, standardized, and objective data from multiple centers is essential for building and training unbiased and generalizable models. Openly available large datasets would facilitate model development and benchmarking.
While these digital innovations are crucial, traditional scientific methods should not be neglected. Large, rigorous clinical trials are required to assess the safety and efficacy of new software. Ideally a competitive market for automated DBS programming software would encourage the continuous improvement and ensure affordable options for patients and healthcare providers.
© Picture: Götz Schleser
