An Overview of the Marie Skłodowska-Curie Doctoral Network “AI in Parkinson’s Disease” (AIPD)
What is Parkinson’s disease?
Parkinson’s disease is an incurable, chronically progressive neurodegenerative disorder and one of the most common neurological diseases worldwide. In Western countries, the number of cases is steadily rising, possibly due to demographic factors. The disease develops over a long period of time, spanning up to 20 years. It is characterized by the death of dopamine-producing cells in the substantia nigra of the brain. The cause of the disease is largely unknown. In addition to the well-known motor symptoms such as tremor, rigidity, and bradykinesia, cognitive, psychiatric, and autonomic symptoms frequently occur. In many cases, these symptoms can only be described qualitatively. Overall, the disease progresses very differently in different patients. Accurate diagnosis and treatment of Parkinson’s disease are difficult. Key challenges are therefore:
- How can we diagnose Parkinson’s disease early, and how can we best treat the symptoms that develop at the right time?
- What can we tell patients about their likely future disease progression?
- What do we do when patients no longer respond to standard treatments?
Data and AI as part of a solution
To address the challenges outlined above, a wide variety of patient-related data is required. This includes, among other things:
- High-dimensional biological omics data, i.e., large biological datasets generated, for example, during the analysis of genes, RNA, proteins, or metabolites
- Medical imaging, specifically MRI scans of the brain
- Data from mobile gait sensors and voice recordings
- Structured longitudinal data from various clinical trials
While traditional statistical methods are typically used to test specific hypotheses for the average patient, these methods often struggle when it comes to making predictions for individual patients based on a wide range of different variables and data modalities. Artificial intelligence (AI) methods offer new possibilities, as they can systematically analyze complex and heterogeneous data and identify patterns that are difficult to detect using conventional methods. This is where the European Marie Skłodowska-Curie PhD Network “Artificial Intelligence in Parkinson’s Disease” (AIPD) comes in. The goal is to systematically integrate modern AI methods with clinical and biological research to create new possibilities for predicting disease risk, subtype, and progression, as well as to personalize and optimize long-term treatment decisions based on data and AI models. All of this is to be carried out on a transparent, ethical, and legally sound basis in the spirit of trustworthy AI.
Key Questions in AI Research on Parkinson’s Disease
AIPD essentially seeks to address three key scientific questions:
How can AI help treat Parkinson’s disease earlier and in a more personalized way? To this end, clinical scores, imaging, omics data, voice recordings, and digital gait sensors are used. Based on this data, individual risk profiles are created, Parkinson’s is distinguished from other conditions (differential diagnosis), disease subtypes are identified, prognostic models are developed, and the response to drug treatments is evaluated. The goal is to develop data-driven models that, in the long term, will be able to support and personalize treatment decisions. Additionally, future clinical trials will be supported in a way that improves the chances of developing new medications for Parkinson’s patients.
How can AI help track the progression of Parkinson’s disease symptoms more objectively using innovative digital technologies? To this end, algorithms are being developed to identify patterns in voice recordings and digital gait sensor data. Based on the accuracy achieved by these models, conclusions can be drawn about the clinical utility of the corresponding digital technologies.
How can trustworthy, explainable AI models be developed that are clinically interpretable as well as legally and ethically sound? Particular emphasis is placed on the explainability and clinical interpretability of AI-based model predictions. Additionally, the generalizability of models is critically examined. From a legal perspective, the interplay with data protection regulations, the AI Regulation, and the Medical Device Regulation is also essential. Finally, a comprehensive ethical assessment of the project is conducted.
Challenges and Potential Solutions from a Data Science Perspective
The development of AI methods to address the questions mentioned above faces a whole range of challenges:
- Data access: Access to patient-related data is highly regulated in Europe. Consequently, a significant amount of effort goes into requesting data access and finalizing legal agreements regarding data use.
- Data harmonization: Medical data from a specific study is generally not representative of the general population. It is therefore essential to evaluate models from different studies to assess their generalizability. However, data from different studies are generally not structurally compatible with one another. They must therefore first be harmonized using a common data model. To this end, agent systems, among other things, are being developed and tested within the framework of AIPD.
- Limited sample sizes: Compared to other fields of application, medical data sets are still relatively small, especially in terms of typical sample sizes in clinical trials. This poses particular challenges for AI methods.
- Short follow-up periods with high heterogeneity: Although clinical trials often run for several years, the number of hospital visits during which data is collected is very limited. In addition, many participants drop out of a study at some point, leading to systematically missing values in the data. At the same time, there is a high degree of heterogeneity in disease progression patterns. The development of AI models for clinical trials is therefore highly complex.
- Multimodal data: Many of the questions mentioned above can only be answered by combining different data modalities. For example, the risk of disease can be derived from a combination of genetic predisposition, age, environmental influences, and lifestyle. This fact makes the development and evaluation of suitable AI models time-consuming and difficult.
- High-dimensional data: Much of the data in medicine is high-dimensional – that is, it has a very large number of features per case – while the number of patients is limited. Examples of this include omics data and medical imaging data. High-dimensional data can very easily lead to overfitting of AI models and thus to poor predictive accuracy.
- Lack of interpretability of sensor data: Data from digital devices, such as smartphones or wearable gait sensors, represent a new and very interesting data source for medicine. However, the raw data is often difficult to interpret at first.
To address the challenges mentioned above, AIPD is adapting, refining, and – in some cases – combining a wide range of data science approaches:
- AI agents: Language models (large language models) and agent systems built upon them are intended to assist in the harmonization of study data.
- Multimodal data fusion: Data from different modalities (e.g., genetic variants, questionnaires on lifestyle or sleep disorders) should be appropriately combined to achieve optimal predictive accuracy.
- Domain adaptation: Models are to be made transferable between different studies based on harmonized data.
- Foundation Models: The foundation models developed in recent years for many data modalities (e.g., spoken language) are compared with established alternatives in terms of predictive accuracy and interpretability.
- Time series models: Many of the research questions outlined above can only be addressed through different modeling approaches for time series. These also include generative approaches, which should be used to simulate possible future disease trajectories at the individual level.
- Hybrid AI: Some of the challenges on the data side can only be addressed by combining different modeling techniques. For example, artificial neural networks are combined with differential equations.
- Causal Machine Learning: Many questions in medicine cannot be solved by mere predictions alone. For example, the question of the potential benefits of a lifestyle change cannot be answered using standard machine learning methods. This is where modern methods from the intersection of AI and statistics come into play.
The Trustworthiness of AI Methods as a Key Consideration
A key consideration in the continued use of AI in medicine is its trustworthiness. This encompasses several aspects addressed in AIPD, such as generalizability, explainability, ethics, as well as data protection and regulation:
- The lack of generalizability of AI models is a major problem in research overall due to the limitations of study data mentioned above. AIPD is therefore fully dedicated to the external validation of models and their domain-specific adaptation based on harmonized data.
- AI models in medicine require careful interpretation to avoid meaningless results and to be able to explain, in the event of potential future use, how a specific result was arrived at.
- The use of highly sensitive data, such as voice recordings, raises ethical questions. Here, a careful balance must be struck against the potential medical benefits. AI systems are subject to a wide range of relevant laws and regulations that must be observed and understood. Furthermore, existing data protection regulations must be adhered to when applying AI algorithms.
Significance for AI Research at Lamarr
AIPD’s interdisciplinary research directly contributes to the core objectives of Lamarr’s Life Sciences & Health division and illustrates how data-driven AI approaches can be applied in medical research: AIPD develops theoretically grounded AI algorithms that harness causal reasoning and critically evaluates foundational AI models. Furthermore, the interpretability of AI models is a central component of AIPD’s research strategy. By utilizing study data from multiple centers, AI approaches can be critically evaluated with regard to their generalizability. Ultimately, AIPD aims to develop robust and trustworthy AI methods that can improve clinical decision-making and support future patient care.
