Paper link: https://www.sciencedirect.com/science/article/pii/S1386505624002168

• Study Design: Individuals with AKI within 48-hour admission were included and temporal laboratory measurements were extracted from two public databases.
• Representation Learning: Laboratory data were split to derive multiple representations with supervised LSTM autoencoders.
• Subphenotypes Derivation: We got the corresponding cluster from multiple representations using K-means. Then, the similarity matrix was derived by computing the frequency of clustering to the same group. Finally, subphenotypes were derived from the similarity matrix by K-means.
• Subphenotype Interpretation: Static patterns analysis, dynamic patterns analysis, and predictive interpretation were conducted to interpret the subphenotypes.

• a. Time series electronic health records (EHR) were utilized as input data for the Bi-LSTM encoder in our study. The Bi-LSTM encoder processed the EHR data, and the hidden state in the final time step was extracted as the representation. Subsequently, the Bi-LSTM decoder employed both the representation and the time-reversed input data to reconstruct the original input data. Simultaneously, the supervisor component was co-trained with the Bi-LSTM autoencoder. This supervisor component utilized the extracted representation as input to predict the probability of mortality [Paper].
• b. We represented EHR in multiple representations in different hidden dimensions. Subsequently, K-means was employed to cluster each representation. In this illustration, we conducted K-means on five representations (left). The similarity of pair-wise patients was quantified as the frequency of being clustered into the same group. For instance, patient 1 and patient 2 were consistently clustered together in four out of five representations, and their similarity was calculated as 4 divided by 5. This similarity matrix, a symmetrical matrix, was constructed to record the pairwise similarities between patients (right).

All the required dependencies are in requirements.txt. For installing all the dependencies run this line of code in command

pip install -r requirements.txt

• Prepare your time-series data into a .pkl file and put it in the folder dataset. You can customize your dataset in utils.py.
• The input data X requires a three-dimension format [number of samples, time dimension, feature dimension].
• The input data Y requires a one-dimension format. The representations will be correlated with data Y by supervised learning [Paper].

To generate representations in multiple dimensions in parallel, run code in command

./01_get_representations.sh

The parameters were illustrated as bellow,

The code results in multiple representations rep_{n}.pkl in the folder ./representations and logs in the folder ./runs. The representations in different hidden dimensions can be visualized as below

• To conduct consensus clustering on representations in multiple dimensions in parallel, run 02_consensus_clustering.py, which will generates the figure of the relative change in area under CDF Curve and the average consensus value of each cluster.

• The derived subphenotypes consensus_cluster_{k}.pkl will be generated in the folder ./results. The consensus matrixes demonstrated the similarity of the pair-wised patients. The blue lines indicated the cluster division.

Further analysis was summarized as below