Download to read offline
Uploaded byDeakin University
Deep learning for episodic interventional data
The document discusses using deep learning models to analyze episodic healthcare data and make predictions. It proposes: 1) Viewing healthcare processes as executable computer programs with hidden "grammars" that can be learned from observational data. 2) Modeling health dynamics as a system of state transitions where treatments shift illness states, and historical events' importance is person-specific. 3) Training models by minimizing prediction loss to forecast outcomes like readmission, mortality, and disease progression based on patients' diseases, treatments, and visits over time.
More Related Content
Similar to Deep learning for episodic interventional data
![CASE_PRESENTATION_ON_subdural_hematoma(SDH)[1 FINAL PPT]-1.pptx](https://cdn.slidesharecdn.com/ss_thumbnails/casepresentationonsubduralhematomasdh1finalppt-1-260129172522-d405d375-thumbnail.jpg?width=640&height=640&fit=bounds)
Deep learning for episodic interventional data
- 1. 28/08/2018 1 Source: rdnconsulting Melbourne, Aug 2018 Truyen Tran Deakin University @truyenoz truyentran.github.io truyen.tran@deakin.edu.au letdataspeak.blogspot.com goo.gl/3jJ1O0 Prognosis Care Deep learning for episodic interventional data
- 2. Setting: Three interactingprocesses in EMR Disease progression Interventions & care processes Recording rules 28/08/2018 2 Source: medicalbillingcodings.org
- 3. Challenge: Complexity Long-term dependencies Irregular timing Mixtureof discrete codes and continuous measures Complex interaction of diseases and care processes Rich domain knowledge & ontologies 28/08/2018 3 visits/admissions time gap ? prediction point Data are not created equally important Models must be accurate AND explainable
- 4. Hypothesis: Healthcare is Turing computational Healthcareprocesses as executable computer program obeying hidden โgrammarsโ The โgrammarsโ are learnable through observational data 28/08/2018 4 https://www.pinterest.com.au/pin/14073817561987932/ http://workingmodeloftheworld.com/Turing-Machine
- 5. Model: Trainable algebraicsystem Health dynamics as a system of transitions of forgettable illness states Treatments โshiftโ illness state from one point to another Medical entities (e.g., disease, treatment, doctor) as algebraic objects (e.g., vector, matrix and function) Importance of historical events are person- specific Training by minimizing prediction loss 28/08/2018 5 Get sick See doctor Enjoy life
- 6.
- 7. Recurrent neural networks Classification Imagecaptioning Sentence classification Neural machine translation Sequence labelling Source: http://karpathy.github.io/assets/rnn/diags.jpeg 28/08/2018 7
- 8. Attention mechanism 28/08/2018 8 Show,Attend and Tell: Neural Image Caption Generation with Visual Attention, K. Xu , J. Ba, R. Kiros, K. Cho, A. Courville, R. Salakhutdinov, R. Zemel, Y. Bengio
- 9. 28/08/2018 9 LSTM LSTMLSTM Readmission Mortality Time-to-event Next diseases Current treatments Diseases Treatments Diseases Treatments Diseases Treatments Visit t1 Visit t2 Visit tN Set function Set interaction Attention Visit embedding Illness stateRESSET #REF: Phuoc Nguyen, Truyen Tran, and Svetha Venkatesh. "Resset: A Recurrent Model for Sequence of Sets with Applications to Electronic Medical Records." IJCNN (2018).
- 10. Health = Recurrence(Set(Illness)โ Set(Intervention)) 28/08/2018 10 illness intervention set function ~ permutation invariant health non-linearity
- 11. Data: Barwon Health,Geelong Australia (2002-2013) 28/08/2018 11
- 12.
- 13.
- 14. Towards a differentiableTuring machine for health 28/08/2018 14 #REF: Hung Le, Truyen Tran, and Svetha Venkatesh. โDual Control Memory Augmented Neural Networks for Treatment Recommendationsโ, PAKDD18. Illness memory Future trajectory Past trajectory
- 15.
- 16. Wrapping up Healthcare canbe viewed as an algebraic system, which can be realized as a learnable Turing program Dynamic diseases-treatments interaction demands new models Deep learning is a viable solution 28/08/2018 162012 Question: where is the likelihood function?
- 17. The Team @Deakin 28/08/2018 17