JMIR Publications

ABSTRACT

Background:

Acute diseases have severe complications that develop rapidly, exhibit distinct phenotypes, and have profound effects on patient outcomes. Predictive analytics can enhance physicians’ care and management of acute disease patients by predicting crucial complication phenotypes for timely diagnoses and treatment. However, ensuring the clinical effectiveness and value for predicting complication phenotypes requires overcoming several challenges. First, patient data collected in early stages (e.g., risk factors) are less informative for predicting phenotypic outcomes. Second, patient data are temporally heterogeneous; e.g., patients receiving laboratory tests at different intervals and with varying frequencies. Third, imbalanced distributions of patient outcomes create additional complexity for complication phenotype predictions.

Objective:

To predict severe complication phenotypes among patients suffering acute diseases, we propose a novel, deep learning–based method that employs recurrent neural network–based sequence embedding to represent patients’ disease progressions, with the consideration of temporal heterogeneities in patient data. The method incorporates a latent regulator to alleviate data insufficiency constraints, by accounting for the dynamics of underlying mechanisms that are not observed in patient data. A cost-sensitive analysis also addresses imbalanced outcome distributions in patient data for better predictions.

Methods:

From a major healthcare organization in Taiwan, we obtain a sample of 10,354 electronic health records that pertain to 6545 peritonitis patients. The proposed method projects these heterogeneous, temporal, clinical data into a substantially reduced feature space, then incorporates a latent regulator (latent parameter matrix) to obviate data insufficiencies and account for variations in phenotypic expressions. Finally, our method applies a cost-sensitive analysis to improve predictive performance further.

Results:

To evaluate the proposed method, we examine its efficacy for predicting two hepatic complication phenotypes for peritonitis patients: acute hepatic encephalopathy (A-HE) and hepatorenal syndrome (HRS). We compare our method with three benchmark techniques. For A-HE predictions, our method attains an area under curve (AUC) of 0.74, which outperforms temporal case-based reasoning (T-MMCBR) by 48%, temporal short long-term memory (T-SLTM) network by 28%, and time fusion convolutional neural network (CNN) by 12%. For HRS predictions, it achieves an AUC of 0.65, which is 20%, 14%, and 18% better than that of T-MMCBR (0.54), T-LSTM (0.57), and time fusion CNN (0.55), respectively. Overall, the proposed method achieves higher recall values for predicting A-HE and HRS than the benchmark techniques.

Conclusions:

The proposed method learns a short-term temporal representation from patient data to predict complication phenotypes, and it offers greater predictive utilities than prevalent data-driven techniques. This method is generalizable and can be applied to different acute disease (illness) scenarios characterized by insufficient patient clinical data availability, temporal heterogeneities, or imbalanced distributions of important patient outcomes.