Note: All contents in this essay are from the tutorial “Time-Series in Healthcare: Challenges and Solutions” at AAAI-22 by Mihaela van der Schaar and Fergus Imrie.
In this essay, I’ll cover some applications for time-series in healthcare, and the challenges as well as solutions for them. Please watch the full video of the tutorial if you are interested in topics that are not covered.
Machine Learning & Medicine/Healthcare/Bio-Science
- ML/AI drives a revolution in medicine
- Medicine drives innovations in ML/AI
Machine learning (ML) and artificial intelligence (AI) are driving a revolution in medicine, and medicine is driving innovations in ML and AI. This kind of relationship encourages improvement in both fields.
Time-Series Models
Time series models are used to forecast events based on verified historical data.
In this tutorial, we focus on machine learning models, some popular ones including RNNs, CNNs, Transformers, and so on.
Below are some useful links that are included in the tutorial.
- Forecasting Big Time Series, Faloutsos et al., KDD Tutorial (2019)
- Understanding LSTM Networks, Christopher Olah
- Gaussian processes for Machine Learning, Rasmussen & Williams
- The Art of Gaussian Processes, Mattos & Tobar, NeurIPS Tutorial (2021)
- Deep Implicit Layers — Neural ODEs, Deep Equilibrium Models, and Beyond, Kolter, Dubenaud & Johnson, NeurIPS Tutorial (2020)
Time-Series in Healthcare: A Multi-Faceted Problem
There are a lot of different kinds of applications for machine learning in time-series in healthcare, but it can be roughly divided into two parts:
- How to apply time series models to healthcare
- How to make time series models as useful as possible
The tutorial will proceed in this order as well.
Part 1: Tailoring Development of Time Series Models to Healthcare Challenges
In the tutorial, multiple challenges were discussed, including:
- Dynamic forecasting
- Time-to-event and survival analysis
- Clustering and phenotyping
- Screening and monitoring
- Early diagnosis
- Treatment effects
To avoid this article being too lengthy, we will focus on the first challenge: dynamic forecasting.
Dynamic Forecasting
One of the most traditional ways of doing dynamic forecasting is using Markov models.
In a Markov model, the disease progression is modeled through passing through a set of disease stages that are observable. However, there are some clear disadvantages to using this kind of model:
- It assumes that all of the disease stages are observable, which is often not the case.
- It models “average” patients and cannot be personalized.
To solve the first problem, hidden Markov models (HMMs) were introduced.
However, some problems still exist. The main ones are:
- It ignores historical data.
- It only captures population-level transitions across progression stages, ignoring individual clinical trajectories.
Central Goals
Before continuing the analysis, let’s introduce the central goals of dynamic forecasting:
- Accurately forecast individual-level disease trajectories
- What are the risks of mortality, relapse, comorbidities, complications, etc. in the future? - Understand disease progression mechanisms.
- Underlying latent structure of disease evolution
- Patients’ subgroup analysis
- Refined phenotypes
Deep Learning Models
Retain [Choi et al., NeurIPS 2016]
To meet these criteria, a number of deep learning models have been developed. One such model is Retain, which is a two-level attention model.
By using attention mechanisms, predictions are interpretable, but the latent structure is still uninterpretable.
Attentive state space models [Alaa & vdS, 2018, NeurIPS 2019]
Another model is attentive state space models (ASSMs). The main idea of ASSMs is to “capture complex, non-stationary representations for patient-level trajectories.”
Simply put, ASSMs maintain the probabilistic structure of HMMs, but use RNNs to model state dynamics.
Attention weights determine the influences of past state realizations on future state transitions.
Part 2: making time series models as useful as possible
Which time-series method to select?
When using machine learning approach to solve problems such as dynamic forecasting, we are facing these problems:
- Which RNN cells to use? (e.g. LSTM, GRU)
- Which architecture to use? (e.g. Bidirectional, Encoder-decoder)
- Attention or not? Long or short memory?
Since manually selecting the best model is impossible due to the time and effort it would take, having a method to do stepwise model selection is very important.
Naively, we can treat performance at each time step as its own black-box function. The objective of the optimization is model performance at each time step. This is a Multi-Objective Bayesian Optimization problem, which tries to find one model with the best trade-off across all objectives. However, this kind of problem is really expensive to compute volume gain with reference to all the objectives.
SMS-DKL [Zhang, Jarrett, vdS, AISTATS 2020]
With its expensive nature, some techniques such as SMS-DKL are introduced. However, it is a little bit complicated and we will skip this part for now.
Conclusion
In this essay, we have discussed some of the challenges and solutions for using time-series models in healthcare. We have also introduced some of the most popular deep learning models for dynamic forecasting.
Time-series analysis is a powerful tool that can be used to improve healthcare outcomes. However, there are still many challenges to overcome before it can be widely adopted. By addressing these challenges, we can make time-series analysis a valuable tool for improving patient care.
References
- Choi, E., Kim, J., Lee, J., Kim, T., & Kim, J. (2016). Retain: A recurrent attention model for interpretable disease progression forecasting. In Advances in Neural Information Processing Systems (pp. 3673–3681).
- Alaa, A., & van der Schaar, M. (2018). Attentive state space models for personalized disease progression forecasting. In Advances in Neural Information Processing Systems (pp. 7788–7797).
- Zhang, J., Jarrett, M., & van der Schaar, M. (2020). SMS-DKL: Scalable multi-objective Bayesian optimization for time series forecasting. In Proceedings of the 37th International Conference on Machine Learning (ICML).
