introductionWith the rapid development of big data and the increasing attention paid by researchers to the medical industry, how to use artificial intelligence to help hospitalized patients early warning, thusImprove survival rate and reduce hospital burdenBecoming increasingly popular.SepsisSepsis

introduction

With the rapid development of big data and the increasing attention paid by researchers to the medical industry, how to use artificial intelligence to help hospitalized patients early warning, thusImprove survival rate and reduce hospital burdenBecoming increasingly popular.

SepsisSepsis.10SepsisICU20%8%Sepsis.Sepsis.

This article uses the MIMIC-IIIv1.4 dataset, a free and publicly available large medical database that covers almost all treatment or diagnostic records required by patients during hospitalization.Integrating comprehensive clinical data from over 40000 unidentified patients admitted to BethIsrael Deaconess Medical Center in Boston, Massachusetts between 2001 and 2012And according to the data usage protocol, international researchers can widely access this data (approximately 1 data point per hour). This database contains 26 tables, and the following will provide a detailed introduction to the dataset.

1. Dictionary Information Auxiliary Table

Dictionary information data, consisting of 5 data tables. When searching for the symptoms or indicators corresponding to a patient's specific code, it is necessary to include them in the dictionary table.The dictionary table makes the structure of the table simpler and clearer.

Introduction to Dictionary Table

2. Patient demography information and hospital turnover information

Patient personal information and main hospitalization information, including 6 data tables. These six tables are more commonly used in conducting experiments.

Introduction to Patient Information Form

Record the patient's personal information. It can be used in conjunction with the ADMISSONS table to briefly analyze the characteristics of admitted patients. In addition, the ADMISSONS table can also serve as a supplement to patient personal information.

..Sepsis.

Usually used to calculate the length of time a patient spends in the ICU.

3. Information table related to outpatient treatment in patient hospitals

Introduction to Medical Record Related Tables

It records which medical services the patient has used that require payment, making it easier to calculate the cost.

Record the patient's ICD-9 diagnostic code, which will be used when studying specific diseases. A patient may have multiple symptoms corresponding to ICD-9 codes, and it is generally believed that the first is the patient's main disease.

LABEVENTS:A record of the patient's laboratory results, such as blood pressure, urine volume, etc. Contains information about laboratory based measurements, it should be noted that the measurement time is the liquid collection time, not the time when clinical staff can use these values.

PRESCRIPTIONSThis table contains order entries related to medication and is of great use when studying patient medication use.

4. Table of treatment related information for patients in the ICU

Patient Treatment Information Record Form

CHARTEVENTSContains all icon records of the patient, displaying their routine vital signs and any other information related to their care, such as ventilator settings, laboratory measurements, mental state, etc.

DATETIMEEVENTSContains all date measurements related to patients in the ICU. It should be noted that in order to protect patient privacy, all dates have undergone privacy processing, but the difference between dates is still meaningful.

NOTEEVENTS..

Sepsis.Divide the patient's historical diagnostic data into multiple channels and input it into the model,And complete the weight calculation of each indicator change and information learning of organ failure in the models of each channel, and finally2/6SepsisThe model framework is shown in the following figure.

Model framework

Multi organ state fusion model.SepsisSepsis.4.

Using GRU as the benchmark model, input different organ failure data into the model in different channelsmultichannelgruSepsis.

Differential temporal weight learning.Due to the dynamic changes in features, it also reflects the patient's physical condition. Therefore, differential temporal weight learning modules were designed in the models of each channel, and attention mechanisms were used to assign corresponding weights based on changes in features, namelyWhile learning about each organ failure, I also learned about the changes in the patient's physical indicators.

Sepsis.Sepsis...

1. Experimental environment configuration

.PyCharmPython3.8Pytorch-gpu1.7

2. Experimental data preprocessing

MIMIC-IIIv1.4...

Data preprocessing process

(1) Marking

MIMIC-IIISepsissepsis3.0.Sepsis2.

.

Oral or injection of antibiotics and body fluid cultures(Blood, urine, cerebrospinal fluid, peritoneum, etc.).7224..

Determination of the time of infection occurrence

SOFA2.Displayed the criteria for determining organ failure.SOFASOFA6.SOFA2.

Sequential Organ Failure Score (SOFA)

Sepsis.SepsisSepsis..2412Sepsis.

Sepsis

Sepsis.2626.

The specific method is:SepsisSepsis26label1Sepsis26label0.

(2) Data aggregation

26.26After aggregation, replace the original indicator with the number of measurements and the maximum value of the indicator during that period.

(3) Routine processing

.261:241:32Extreme sample imbalance.1:2.

2MIMIC-IIIv1.4.2070%.

Data missing rate

knn.knn.

.scaler.

.

Dataset Statistics

3. Experimental results

.

Complete experimental results

lrepoch.

.

(1) Analysis of benchmark model experimental results

MIMIC-IIISepsisThe benchmark model only covers the following two methodsSepsis.

XGBoostJeromeFriedman,2000

LightGBMGuolinKe,2017

GRUJunyoungChung,2014

Using the two machine learning algorithms mentioned above as a referenceBenchmark model for GRU comparisonGRU.

.

Gru and benchmark model experimental results

aucROC0.75.Neural networks consider the interrelationships of time seriesaucf1f120.gru.

In addition, in order to verify the proposed approach in this articleThe Effectiveness of Differential Timing Based on Attention Mechanism.

Delta_ Gru:attention_gru.delta.

.

Experimental results of differential timing with different implementation methods

The table above shows thatgrudelta.Utilizing differential timing with attention mechanism is more effective than simply expanding the original dataset features.

Due to the weight calculation of attention mechanism,Have a corresponding importance selection for the dynamic change range of the original feature.

Experimental results of differential timing under multi organ state fusion

The table above shows thatMulti organ state fusion modeldelta.gru.

(2) Results of ablation experiment

The previous benchmark model is to prove the necessity of the model or module, whileThe ablation experiment is to demonstrate the effectiveness of each module in the model...

Experimental results of multi organ state fusion ablation

.gru166.

Such a significant improvement is very rare and surprising for the model itself, indicating that this article.

Experimental results of multi organ state fusion ablation

.On the basis of differential timing, the effect of the model is significantly improved by incorporating multi organ state fusion design.

auc14f176..

Results of Differential Sequential Ablation Experiment

The above table shows that in the gru model,The Importance of Differential Timing Module.auc7.

.

(3) Model robustness results

..

All experimental results areMIMIC-IIIv1.4 dataset based on 2-hour aggregation.

Experimental results of 2-hour aggregated dataset

6MIMIC-IIIv1.4.

Experimental results of 6-hour aggregated dataset

26.

summary

.Sepsis26.

Sepsisauc0.8-0.90.80.9.MIMIC-IIIv1.4.

Therefore, in order toSepsisSepsisSepsisSepsis.

30..

Disclaimer: The content of this article is sourced from the internet. The copyright of the text, images, and other materials belongs to the original author. The platform reprints the materials for the purpose of conveying more information. The content of the article is for reference and learning only, and should not be used for commercial purposes. If it infringes on your legitimate rights and interests, please contact us promptly and we will handle it as soon as possible! We respect copyright and are committed to protecting it. Thank you for sharing.(Email:[email protected])