Learning Disease Progression Models That Capture Health Disparities

The model makes assumptions upon the expression of disparities while being identifiable. The underlying process must verify the assumptions. It would be beneficial to discuss further how realistic and/or common these assumptions are. An analysis of a misspecified model when the underlying generating process does not meet these assumptions would be valuable.

• While the model is interesting from a clinical perspective, I am not sure this is the right venue for this publication due to limited technical novelty
• Very limiting disease progression trajectory due to linear assumption in time. A patient can therefore not experience worsening and improvement on their disease trajectory
• The proofs show that not taking into account disparities will bias the result, however other (non-linear) disease progression models can take "baseline" covariates like ancestry, effectively conditioning the probability of the latent progression $z$ on $A$ as well. Comparisons to models like these are necessary to underline the claims.
• The authors need to compare the disease progression modeling with other state of the art methods of inferring latent disease trajectories. It is not clear how well more realistic disease trajectories are captured and how the model compares to SOTA models on this task.
• For the synthetic data it would be helpful to also use data not sampled from the same architecture as the model under consideration, but more complex disease progressions to examine how well the model can infer these despite the relatively simple assumptions

However, the paper suffers from several flaws. I am really willing to increase my grade if those points are addressed, but as it is, the contribution of the model compared to existing strategies, in terms of performance is really unclear.

• the baselines seem quite weak. The authors report that several indicators are important, like the visit frequency. I am not sure to understand from the manuscript which features the baselines include, in particular do they include demographics information? and not sure either that PCA or FA are the best tools to do feature engineering. In particular all those approaches are linear, so they can not model interactions between demographics and other features. Maybe considering tree-based methods would be relevant here, with an appropriate feature engineering process. I know that prediction is not the end-goal of the model, but this constitutes the only "measurable" performance indicator.
• the ablation studies are interesting, but they should also be conducted on the real-data application, to assess the variation in predictive performance, because it is not very convincing that data simulated with a model would be less accurately represented by a different model.
• additional baselines to consider, both on simulated and real data it would be interesting to compare the performances with models that account for demographics characteristics differently (with one type of disparity for instance)

1. The proposed method appears to be a variant of a hidden Markov model (HMM). Instead of using transition probabilities in HMM, it employs simple functions to describe transitions between states and outcomes. This simplification might limit the model's ability to capture the complex dynamics of disease progression.
2. The directed acyclic graph, the selection of functions between observed and hidden variables, and the specific types of disparities incorporated in Section 3 seem overly simplistic and heuristic. The paper lacks detailed reasoning or motivation for these design choices. Providing justifications or empirical evidence supporting these decisions would enhance the credibility of the model.
3. The paper adopts a linear representation solely because "it provides an interpretable characterization of the trajectory." However, real-world disease progression often involves intricate, nonlinear relationships between variables. Relying solely on a linear model may lead to suboptimal performance and may not capture the true underlying patterns, potentially undermining the trustworthiness of the explanations. Exploring nonlinear models could offer better performance and more reliable interpretations.
4. The empirical results are based on a private dataset with a relatively small number of subjects (n=2,942) and only four features. This raises concerns about the model's generalizability to other datasets. I recommend validating the model on public EHR datasets such as MIMIC or UK Biobank to assess its broader applicability. Furthermore, the comparisons are limited to simple baselines like linear regression, quadratic regression, principal component analysis, and factor analysis. Evaluating the model against state-of-the-art methods, including neural networks, RNNs, and transformers, would provide a more comprehensive assessment.
5. Insufficient Evidence of Superior Performance:
   • The model's explainability is only qualitatively assessed using medical knowledge. Incorporating quantitative evaluations or user studies could strengthen the claims about interpretability.
   • The model does not show (significant) improvements over the baselines in reconstruction and predictive tasks. Additionally, the absence of error bars or statistical significance tests makes it challenging to determine if the differences are meaningful. If the authors can include the code, my confidence in the results will be higher.