Sysrev automatically builds machine learning models at every stage of  review.

In each sysrev users can screen articles in their corpus by marking them as an 'include' or 'exclude'.  While reviewing a sysrev screening model is silently learning how to replicate reviewer decisions.  Screening models can help accelerate the review process and eventually automate reviews.  Automated reviews will be a key step in the updateable 'living' reviews talked that will be the next frontier for document review [1,2].  

Sysrev screening models take the modeling process one step further by using them to drive the order of document review.  The process is simple - humans label articles, models learn from labeled examples and articles are re-ordered to maximize their learning value.

A machine learning process prioritizes review in every sysrev.

The current performance of the document screening models can be evaluated on the existing human labels at any time in the overview via prediction histograms. Public reviews like the ERAS projects show strong performance of screening models.  

Automated screening was the first sysrev machine learning tool. This algorithm quietly prioritizes human screening to maximize the performance of a trained screening model. Resulting models sometimes work very well! The above three models come from some of the completed ERAS Spine and Craniotomy projects. The x-axis gives the screening models probability of inclusion. Green bars are the number of articles that were actually included by a reviewer and red bars the number of articles actually excluded by a reviewer. These models work pretty well, but are not perfect.

How do the screening models actually work?

This is an active area of development for sysrev.com.  Currently, sysrev models are rebuilt whenever reviewers complete 25 articles.   The screening models use a combination of word-embedding models and multi-task neural networks. The stages are: 1. document featurization 2. pre-processing 3. multitask learning.

Document Featurization: A word2vec and paragraph2vec model is built from pubmed and wikipedia and other document stores.  These are embedding models that transform words/paragraphs into numeric feature vectors built from the co-occurrence of nearby words in the pubmed corpus.  Sysrev uses the deeplearning4j package for both models (see word2vec and doc2vec documentation) [3,4].

Pre-processing: The modeling container first performs a resampling procedure to create a more balanced training set for the following supervised learning step.  A few procedures are at use right now, but the simplest resamples the minority class documents until they have equal prevalence with the majority class.  Future directions involve generating synthetic training data.

Multitask Learning: After resamping and featurization a multitask feed forward neural network is built.  The outputs are the label values assigned by reviewers. The below simple image describes the shallow, feed-forward, multitask model built for a prostate cancer review at sysrev.com/p/844.  In short, all the documents in the review are featurized and fed through two hidden layers, a batch normalization layer and end in many cross-entropy softmax outputs. Currently these models only handle binary and categorical labels.  To learn more about the benefits of multitask learning on problems like these I strongly recommend Sebastian Ruder's blog [5].

A simple shallow, multitask neural network is trained whenever reviewers complete 25 article reviews. The above diagram is a truncated example for a review of natural products in prostate cancer at sysrev.com/p/844 .

Future Directions

Subscribe to be notified when we complete these next steps.  Your subscription encourages us to keep writing!

We have many new innovations coming in this year.  Among other machine learning goals, we plan to implement: LSTM classifiers,  shared labeling, and concordance driven review.

Sysrev label classifiers will move from shallow multitask networks to an LSTM neural network architecture this year.

Shared labeling will allow labels to be used across sysrevs. A common label we have seen in multiple projects is 'species'. It is possible to automate this label, and (if shared) this label would not have to be retrained for every new project.

Communication between human brains is notoriously difficult.  You create a review, recruit friends, and define some tasks.  Unfortunately, each friend brings their own biases and misunderstandings.  Sysrev's are an excellent place to start picking apart human concordance. Our analytics platform helps reviewers manage concordance [6].  Soon concordance will be an indispensable in every multi-person review.

References

[1] Pharmacovigilance -- "Living Reviews" Applied https://blog.sysrev.com/pharmacovigilance-living-reviews/

[2] Out of date before it's published https://www.natureindex.com/news-blog/living-systematic-reviews-emerging-solution-problem-superseded-research-zika-virus

[3] Word2Vec, Doc2vec & GloVe: Neural Word Embeddings for Natural Language Processing

[4] Doc2Vec, or Paragraph Vectors, in Deeplearning4j https://deeplearning4j.org/docs/latest/deeplearning4j-nlp-doc2vec

[5]Sebastian Ruder (2017). An Overview of Multi-Task Learning in Deep Neural Networks. arXiv preprint arXiv:1706.05098. ruder.io/multi-task/index.html#whatshouldishareinmymodel

[6] A New Analytics Dashboard! https://blog.sysrev.com/user-inclusion-concordance/