The first stage of this project has been focused on defining the problem rather than building a final model. Alzheimer’s disease is studied through a wide range of measurements, including cognitive assessments, neuroimaging, blood-based biomarkers, genetic information, and clinical variables. These measurements provide different types of information about an individual, but they are not necessarily collected at the same time, with the same frequency, or with the same level of completeness.
This makes longitudinal modelling particularly challenging. A participant may have several observations for one modality and very few for another, while the intervals between visits can also vary considerably. As a result, the information available to a model is not fixed. The relative usefulness of a measurement may also depend on the participant’s current state and on what has been observed previously.
The main research question I am investigating is whether a multimodal model can account for these differences instead of treating all available information as equally informative. In particular, I am interested in whether the contribution of individual modalities can be adjusted according to their availability, estimated reliability, and behaviour over time. This also raises a related question about missing information: rather than treating missing measurements only as a technical problem, can their absence be represented explicitly as part of the modelling process?
The initial framework has therefore been designed around longitudinal and multimodal data rather than around a single prediction task. It includes representations of individual trajectories, changes from baseline, rates of change, and other characteristics of temporal behaviour. Different modelling configurations can then be compared to determine whether adaptive multimodal integration provides an empirical advantage over simpler approaches.
At this stage, I am deliberately not treating the proposed approach as an established improvement. The central idea is a hypothesis that needs to be tested. If adaptive weighting does not improve performance, calibration, robustness, or another predefined measure of model behaviour, that result is just as important to the research as an improvement would be. For the same reason, individual components of the framework will need to be evaluated separately rather than being assumed to be necessary simply because they are part of the proposed architecture.
The current implementation provides the foundation for those experiments. It includes multimodal representation, longitudinal analysis, model comparison, missing-modality testing, robustness experiments, uncertainty estimation, calibration analysis, and ablation procedures. These components are being developed as parts of an experimental framework, allowing the underlying research question to be examined from several directions rather than relying on a single performance measure.
An important distinction has also been established at this stage between the longitudinal research framework and the separate biomarker dataset included in the implementation. The latter contains 2,500 synthetic participants and represents a cross-sectional dataset rather than a longitudinal patient cohort. It is therefore being kept separate from the longitudinal analyses and will be used for biomarker-based classification experiments rather than presented as evidence from real clinical participants.
For now, the project remains at the stage of establishing the research framework and defining how the central hypothesis can be tested rigorously. The objective is not to assume that a more complicated model will necessarily be a better one, but to determine through controlled comparisons whether the additional modelling of modality reliability, missing information, and temporal structure contributes meaningful improvements.