Temporal Phenotyping using Deep Predictive Clustering of Disease Progression

Changhee Lee, Mihaela van der Schaar

Presented at ICML 2020, this paper introduces a novel deep learning framework for "temporal phenotyping"—the process of grouping patients based on how their diseases progress over time. Unlike traditional clustering, which only looks at past data, this method ensures that patients in the same group share similar future clinical outcomes.

TL;DR

The authors propose AC-TPC (Actor-Critic Temporal Predictive Clustering), a deep learning model that clusters time-series patient data based on predicted future outcomes. By using an actor-critic framework to handle discrete cluster assignments, the model identifies clinically meaningful patient subgroups that offer significantly better prognostic value than traditional unsupervised methods.

What Problem Does This Paper Address?

In medicine, "phenotyping" is the act of categorizing patients into subgroups to better tailor treatments. With the rise of Electronic Health Records (EHRs), researchers have increasingly used machine learning to cluster patient time-series data. However, most existing methods are purely unsupervised: they group patients based on similarities in their past measurements (e.g., similar heart rate trends).

The problem is that two patients who look similar today might have vastly different futures—one might recover while the other develops a severe comorbidity. Traditional clustering is "outcome-blind," leading to heterogeneous groups that lack prognostic utility.

Temporal Phenotyping Concept

The authors aim to bridge this gap by creating predictive clusters. As illustrated above, the defining characteristic of a group in this framework is a shared future trajectory, ensuring that patients clustered together are expected to experience similar future clinical events.

Methods

The authors introduce AC-TPC, which consists of three main neural network components and a set of learnable cluster centroids:

AC-TPC Architecture
  1. The Encoder: A Recurrent Neural Network (specifically an LSTM) that compresses a patient's variable-length history into a continuous latent representation ($z_t$).
  2. The Selector (The Actor): A fully-connected network that takes the latent representation and assigns the patient to one of $K$ clusters. Because picking a cluster is a discrete, non-differentiable action, the authors use Actor-Critic reinforcement learning to allow the model to "backpropagate" through the selection process.
  3. The Predictor (The Critic): This network estimates the probability distribution of future outcomes. It is trained to predict outcomes based on either the specific patient's encoding or the assigned cluster's centroid.
  4. The Embedding Dictionary: A set of centroids in the latent space. Each centroid represents the "typical" future outcome distribution for its cluster.

The Secret Sauce: The model minimizes the Kullback-Leibler (KL) divergence between the outcome distribution predicted for an individual patient and the distribution associated with their assigned cluster centroid. This forces the clusters to be "outcome-homogeneous."

Datasets

The model was evaluated on two significant real-world medical datasets:

  • UK Cystic Fibrosis Registry (UKCF): Data from 5,171 patients with 89 variables (demographics, lung function, infections). The goal was to cluster patients based on the development of 22 different comorbidities in the following year.
  • Alzheimer’s Disease Neuroimaging Initiative (ADNI): Data from 1,346 patients tracking cognitive tests and biomarkers. The goal was to cluster patients based on their progression between three diagnostic stages: normal, mild cognitive impairment, and Alzheimer’s.

Key Insights and Ideas

  • Prognostic Homogeneity: The core "aha moment" is that a cluster is only useful to a clinician if everyone in that cluster has a similar prognosis. AC-TPC prioritizes this by making the clustering objective explicitly dependent on future labels.
  • Dynamic Phenotyping: Because the model processes data at each time step, a patient’s "phenotype" can change. For example, if a patient with diabetes suddenly shows signs of asthma, the model can dynamically reassign them to a "Diabetes + High Risk of Asthma" cluster.
  • Handling High-Dimensional Outcomes: In the UKCF dataset, there are $2^{22}$ possible combinations of future comorbidities. AC-TPC successfully finds a small number of "latent" clusters (e.g., 12) that capture the most common and clinically relevant combinations, making the results interpretable for doctors.
  • Stability Against Over-Clustering: Unlike $K$-means, which often struggles if you pick the "wrong" $K$, AC-TPC’s loss functions (specifically the embedding separation loss) help it ignore redundant clusters, effectively "activating" only the number of clusters actually needed to describe the data.

Key Results

AC-TPC consistently outperformed state-of-the-art benchmarks like SOM-VAE and Deep Clustering Network (DCN) extensions:

  • Clustering Quality: On the UKCF dataset, AC-TPC achieved an Adjusted Rand Index (ARI) of 0.602, nearly doubling the performance of the next best predictive benchmark (KM-E2P at 0.440).
  • Predictive Accuracy: It achieved an AUROC of 0.843 for comorbidity prediction, showing that the clustering process actually helps, rather than hinders, the ability to forecast the future.
  • Clinical Relevance: The authors demonstrated that the model could identify specific clusters with high risks for certain comorbidities (e.g., a cluster where 89% of patients developed asthma vs. <3% in others), providing actionable information for preventative care.

Limitations and Future Work

  • Interpretability of Latent Space: While the centroids help, the underlying encoder is still a "black box" LSTM. Future work could focus on making the relationship between specific input features and cluster assignment more transparent.
  • Data Requirements: Like most deep learning models, AC-TPC requires a large amount of longitudinal data with labeled outcomes to train effectively.
  • Fixed $K$ Upper Bound: While the model can "deactivate" clusters, the user still needs to define a maximum number of potential clusters ($K$) at the start.

Overall, AC-TPC represents a significant step toward "precision medicine" by providing a framework that groups patients not just by who they are, but by where their health is headed.

Summary updated

Abstract

Due to the wider availability of modern electronic health records, patient care data is often being stored in the form of time-series. Clustering such time-series data is crucial for patient phenot...
BibTeX
@inproceedings{lee2020temporal,
  author = {Changhee Lee and Mihaela van der Schaar},
  title = {{Temporal Phenotyping using Deep Predictive Clustering of Disease Progression}},
  booktitle = {International Conference on Machine Learning},
  year = {2020},
  url = {https://openreview.net/forum?id=6ELJ7jNDPD},
}
Figure