文献阅读记录

0. Title

An Interpretable ICU Mortality Prediction Model Based on Logistic
Regression and Recurrent Neural Networks with LSTM units.

1. Research Objective

• To propose a model that combines sequential features that include multiple values over the course of the patient’s ICU stay (e.g. a sequence of pulse measurements) and non-sequential features that only have a single value in the dataset, such as diagnoses or procedures recorded prior to the prediction time point.
• Improving the interpretability of RNNs with LSTM units by providing the top ten positive or negative features and the corresponding weights to indicate significances of different features.

2. Background and Problems

• Although the existing studies utilizing RNNs achieved higher accuracy of ICU mortality prediction compared with the traditional scoring systems based on logistic regression, they cannot provide explicit interpretability as scoring system can, and therefore lack face validity.
• Additionally, most of the existing studies on RNN mainly considered the sequential features (e.g. vital signs or laboratory test results) extracted from EMR data, but did not describe methods for combined analyses of non-sequential (e.g. patient demographics, diagnoses and procedures) features together with sequential features.

3. Method

The propsed model

最终的结果由非时间序列模型和LSTM学到的各个时间序列的表示作为输入来进行预测。

作者认为其中时间序列数据和非时间序列的权重向量可以用来表示各个特征的重要性。

• Rank the features based on these weights in order to find out the top ten positive features and the top ten negative features that have contributed most strongly to the prediction of ICU mortality.

Data

The features we used in the ICU mortality prediction model include both sequential features and non-sequential features. The non-sequential features include patient demographics, diagnoses, medications, Braden score (risk of development of decubitus ulcers), blood transfusions and procedures. The sequential features include vital signs, Glasgow Coma Scale (GCS), and laboratory test results.

In the RNN-LSTM model, each of the sequential features was represented by the sequences of 48 values corresponding to the 48 hours in the first two calendar days after ICU admission. If there was more than one value available during a particular hour, the mean of the values during that hour was calculated. If there was no value reported during a particular hour, missing value was set. We used mean and standard deviation to transform real values into categorical values; missing values were assigned to a special category. In the comparison logistic regression model, each sequential feature was represented by two variables: a) the mean of the values reported during the first two calendar days after ICU admission and b) the slope of a line fitted through the values reported during the first two calendar days after ICU admission using least squares. In this way, both RNN-LSTM and logistic regression model represented the trends of sequential features.

4. Evaluation

5. Conclusion

6. Notes

References

Che Z, Purushotham S, Cho K, Sontag D, Liu Y. Recurrent neural networks for multivariate time series with missing values. arXiv preprint arXiv:1606.01865. 2016.

这篇文章针对RNN处理多元时间序列进行研究，通过引入二进制掩码可以得到关于缺失值的模式。