Healthcare as an algebraic system

Throughout the history of homo sapiens, wars, famine and diseases are the main killing machines.  For the first time, we have recently defeated famine and lost appetite for wars. We now turn our attention to the most basic need of all: health. Who doesn't want to live a healthy life and die in peace?

But how do understand a healthcare system, from a modeler point of view? Healthcare is a complex business. A battle field that can determine outcome of election, which may change the course of history.

I've always speculated that healthcare is an algebraic system. Medical objects such as disease, treatment, medication, theater, department and even doctor can be represented as algebraic objects. Processes, protocols, rules and the like can be defined as algebraic operators.

One simplest object representation is vector. And this is very powerful, given the fact that most data manipulation machineries are developed for vectors. Matrices, tensors, graphs and other more sophisticated representations are much less developed.

The algebraic view of things in the world has been there for long. Recently it comes in "embedding" of the objects, from word to sentence to document to graph to anything2vec.

In our recent work "Finding Algebraic Structure of Care in Time: A Deep Learning Approach", we take this view to its full. Diseases and treatments are embedded into vectors. A hospital visit is also a vector, which is computed as a difference of illness vector and treatment vector. This gives rise to simple modelling of disease-treatment motifs, and visit-visit transitions. The entire temporal progression can be modelled as recurrence of simple linear matrix-vector multiplications, or a more sophisticated LSTM.

Once the health state at each visit can be determined, decision making can be acted upon, based on the entire trajectory, with attention to the recent events, or in absence of any time-specific knowledge, using content-based addressing.

After all, healthcare is like a computer execution of a health program, which is jointly determined by the three processes: the illness, the care and the recording. It is time for a Turing machine.

Stay tuned. The future is so bright that you need to wear sun glasses.

Our recent contributions

• Dual memory neural computer for asynchronous two-view sequential learning, H Le, T Tran, S Venkatesh, KDD'18
• Dual control memory augmented neural networks for treatment recommendations, H Le, T Tran, S Venkatesh, PAKDD'18
• Resset: A recurrent model for sequence of sets with applications to electronic medical records, P Nguyen, T Tran, S Venkatesh, IJCNN'18.
• Graph Memory Networks for Molecular Activity Prediction, Trang Pham, Truyen Tran, Svetha Venkatesh, ICPR'18
• Finding Algebraic Structure of Care in Time: A Deep Learning Approach, Phuoc Nguyen, Truyen Tran, Svetha Venkatesh, NIPS Workshop on Machine Learning for Health (ML4H), 2017.
• Deep Learning to Attend to Risk in ICU, Phuoc Nguyen, Truyen Tran, Svetha Venkatesh, IJCAI'17 Workshop on Knowledge Discovery in Healthcare II: Towards Learning Healthcare Systems (KDH 2017).
• Graph Classification via Deep Learning with Virtual Nodes Trang Pham, Truyen Tran, Hoa Dam, Svetha Venkatesh, Third Representation Learning for Graphs Workshop (ReLiG 2017).
• Learning Recurrent Matrix Representation, Kien Do, Truyen Tran, Svetha Venkatesh.Third Representation Learning for Graphs Workshop (ReLiG 2017), also: arXiv preprint arXiv: 1703.01454.
• Predicting healthcare trajectories from medical records: A deep learning approach,Trang Pham, Truyen Tran, Dinh Phung, Svetha Venkatesh, Journal of Biomedical Informatics, April 2017, DOI: 10.1016/j.jbi.2017.04.001.
• Deepr: A Convolutional Net for Medical Records, Phuoc Nguyen, Truyen Tran, Nilmini Wickramasinghe, Svetha Venkatesh, IEEE Journal of Biomedical and Health Informatics, vol. 21, no. 1, pp. 22–30, Jan. 2017, Doi: 10.1109/JBHI.2016.2633963.
• Stabilizing Linear Prediction Models using Autoencoder, Shivapratap Gopakumara, Truyen Tran, Dinh Phung, Svetha Venkatesh, International Conference on Advanced Data Mining and Applications (ADMA 2016).
• DeepCare: A Deep Dynamic Memory Model for Predictive Medicine, Trang Pham, Truyen Tran, Dinh Phung, Svetha Venkatesh, PAKDD'16, Auckland, NZ, April 2016. 
• Learning vector representation of medical objects via EMR-driven nonnegative restricted Boltzmann machines (e-NRBM), Truyen Tran, Tu D. Nguyen, D. Phung, and S. Venkatesh, Journal of Biomedical Informatics, 2015, doi:10.1016/j.jbi.2015.01.012.
• Tensor-variate Restricted Boltzmann Machines, Tu D. Nguyen, Truyen Tran, D. Phung, and S. Venkatesh, AAAI 2015.
• Latent patient profile modelling and applications with Mixed-Variate Restricted Boltzmann Machine, Tu D. Nguyen, Truyen Tran, D. Phung, and S. Venkatesh,  In Proc. of 17th Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD’13), Gold Coast, Australia, April 2013.