Background
-
We present a deep learning model to utilize dynamic treatment information for predicting MACE after ACS, and the incorporation of dynamic treatment information into learning boosts the performance of MACE prediction.
-
The proposed model extracts the latent representation of dynamic treatment features via Bi-LSTM, which can be used to predict whether a patient occurs MACE in his or her hospitalization.
-
Extensive experiments are conducted on a real EHR dataset, which consists of 2930 ACS patient samples collected from a Chinese hospital, to demonstrate the effectiveness of our proposed model for MACE prediction.
Related work
Methods
Patient feature processing and embedding
Using Bi-LSTM to generate deep representations of treatment information
MACE prediction
Results
Data collection
Characteristics | No. of participants (n = 2930) |
---|---|
Demographics | |
Age, (mean ± sdv.) [min-max] | 62.27 ± 12.11 [28–91] |
Gender, Male/Female | 2079/851 |
Physical examination, (mean ± sdv.) [min-max] | |
Systolic BP, mm Hg | 132.10 ± 17.64 [11–240] |
Diastolic BP, mm Hg | 77.59 ± 10.17 [35–120] |
Height, cm | 167.01 ± 8.15 [56–188] |
Weight, kg | 71.81 ± 12.42 [32–200] |
Ejection Fraction, (mean ± sdv.) [min-max] | 59.51 ± 7.82 [17–78] |
Comorbid conditions (%) | |
Diabetes | 803 (27.4%) |
Hypertension | 1981 (67.6%) |
Heart Failure | 165 (5.6%) |
arteriosclerosis | 2267 (77.4%) |
History of current or previous smoking | 1113 (38.0%) |
Laboratory data, (mean ± sdv.) [min-max] | |
Creatinine, umol/L | 78.72 ± 38.18 [29.5–739.4] |
Creatinine kinase, umol/L | 86.11 ± 112.02 [6.2–4651.1] |
Alanine aminotransferase, umol/L | 26.02 ± 27.66 [1.7–593] |
Aspartate aminotransferase, umol/L | 21.70 ± 19.53 [5.8–589.4] |
Troponin T, ng/ml | 0.029 ± 0.084 [0.002–0.886] |
Glucose, umol/L | 6.16 ± 2.26 [2.69–28.62] |
Disease/Treatment history (%) | |
Post-PCI (patient who has taken PCI surgery in the past and was admitted into the hospital at this time) | 816 (27.8%) |
Post-CABG (patient who has taken CABG surgery in the past and was admitted into the hospital at this time) | 46 (1.6%) |
Length of Stay, (mean ± sdv.) [min-max] | 9.12 ± 7.05 [1–54] |
MACE (%) | 752 (22.4%) |
Experiment settings and baseline models
Data analysis
AUC | Accuracy | |
---|---|---|
LR | 0.637 ± 0.010 | 0.752 ± 0.007 |
Mix | 0.681 ± 0.006 | 0.746 ± 0.005 |
Dynamic |
0.713 ± 0.005
|
0.764 ± 0.004
|
Boosted-RMTM | 0.700 ± 0.003 | 0.689 ± 0.004 |
Effect of length of stay
Effect of training set size
Statistical test
Model | Boosted-RMTM | LR | Dynamic | Mix |
---|---|---|---|---|
Boosted-RMTM | / | 5.98E-5** | 5.46E-11** | 0.018* |
LR | / | / | 4.68E-7** | 1.17E-3** |
Dynamic | / | / | / | 1.16E-9** |
Mix | / | / | / | / |
Discussions
-
In most cases, our proposed dynamic model outperforms benchmark models for predicting MACE after ACS. The p-values between the proposed model and benchmark models show that there is a significant difference between the performances obtained by the employed models. Our proposed dynamic model has an average AUC of 0.713 and is thus the best MACE predictor. These findings confirm our assumption that leveraging dynamic treatment information contained in a large volume of heterogeneous EHR appears to boost the performance of MACE prediction, and has significant potential to meet the demand clinical prediction of other diseases, from a large volume of EHR in an open-ended fashion.
-
With the gradual inclusion of more treatment information into learning for individuals, the prediction performance dramatically increases. The tendency of the curve in Fig. 5 arises as the hospitalization day per patient increases. As well, it is clearly to see that the curve surpasses 0.7 in terms of AUC after the number of hospitalization days is larger than five. It indicates that we need at least 5 days’ treatment information per patient to obtain the stable prediction results.
-
It is not surprising to see from Fig. 6 that with sufficient training data samples, the proposed model can achieve a better prediction performance since deep learning method can achieve accurate representations from the big data. Due to the large amount of EHR data generated over time, we plan to investigate the suitability of deep neural networks for discovering nontrivial knowledge that best describe the inpatient treatment journeys and then improve the performance of MACE prediction.
Limitations
-
For one thing, the dynamic nature of patient status is often essential/critical to the selection of treatment interventions. To address this challenge, we expect that our proposed model can incorporate richer execution information, e.g., vital signs, symptoms, and clinical observations on patient status, etc., into learning, which would make our proposed model more intelligent in the treatment adoption and MACE prevention.
-
For the other thing, our proposed model neglects the causal relations between treatment interventions and effects. Note that causal effect analysis is useful to find out unexpected changes in treatment interventions and explain why scheduled treatment plans are changed to obtain the optimal treatment effects. As an open medical problem, the causal effect analysis can be benefited in mining a large scale of EHR data in a maximum-informative manner.