Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2016 Aug:56:301-318.
Epub 2016 Dec 10.

Doctor AI: Predicting Clinical Events via Recurrent Neural Networks

Edward Choi  1 , Mohammad Taha Bahadori  1 , Andy Schuetz  2 , Walter F Stewart  2 , Jimeng Sun  1
Affiliations

Doctor AI: Predicting Clinical Events via Recurrent Neural Networks

Edward Choi et al. JMLR Workshop Conf Proc. 2016 Aug.

Abstract

Leveraging large historical data in electronic health record (EHR), we developed Doctor AI, a generic predictive model that covers observed medical conditions and medication uses. Doctor AI is a temporal model using recurrent neural networks (RNN) and was developed and applied to longitudinal time stamped EHR data from 260K patients over 8 years. Encounter records (e.g. diagnosis codes, medication codes or procedure codes) were input to RNN to predict (all) the diagnosis and medication categories for a subsequent visit. Doctor AI assesses the history of patients to make multilabel predictions (one label for each diagnosis or medication category). Based on separate blind test set evaluation, Doctor AI can perform differential diagnosis with up to 79% recall@30, significantly higher than several baselines. Moreover, we demonstrate great generalizability of Doctor AI by adapting the resulting models from one institution to another without losing substantial accuracy.

PubMed Disclaimer

Figures

Figure 4
Figure 4
Architecture of GRU
Figure 1
Figure 1
This diagram shows how we have applied RNNs to solve the problem of forecasting of next visits’ time and the codes assigned during each visit. The first layer simply embeds the high-dimensional input vectors in a lower dimensional space. The next layers are the recurrent units (here two layers), which learn the status of the patient at each timestamp as a real-valued vector. Given the status vector, we use two dense layers to generate the codes observed in the next timestamp and the duration until next visit.
Figure 2
Figure 2
Characterizing behavior of the trained network: (a) Prediction performance of Doctor AI as it sees a longer history of the patients. (b) Change in the perplexity of response to a frequent code (hypertension) and an infrequent code (Klinefelter’s syndrome).
Figure 3
Figure 3
The impact of pre-training on improving the performance on smaller datasets. In the first experiment, we first train the model on a small dataset (red curve). In the second experiment, we pre-train the model on our large dataset and use it for initializing the training of the smaller dataset. This procedure results in more than 10% improvement in the performance.
See this image and copyright information in PMC

References

    1. Bahadori Mohammad Taha, Liu Yan, Xing Eric P. Fast structure learning in generalized stochastic processes with latent factors. KDD. 2013:284–292.
    1. Bastien Frédéric, Lamblin Pascal, Pascanu Razvan, Bergstra James, Goodfellow Ian J., Bergeron Arnaud, Bouchard Nicolas, Bengio Yoshua. Theano: new features and speed improvements. Deep Learning and Unsupervised Feature Learning NIPS 2012 Workshop. 2012
    1. Bengio Yoshua. Deep learning of representations for unsupervised and transfer learning. Unsupervised and Transfer Learning Challenges in Machine Learning. 2012;7:19.
    1. Bengio Yoshua, Courville Aaron, Vincent Pierre. Representation learning: A review and new perspectives. Pattern Analysis and Machine Intelligence, IEEE Transactions on. 2013;35(8):1798–1828. - PubMed
    1. Che Zhengping, Kale David, Li Wenzhe, Bahadori Mohammad Taha, Liu Yan. Deep computational phenotyping. Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; ACM; 2015. pp. 507–516.

LinkOut - more resources