Microsoft Word - Deep Learning Vs Traditional Models_Abdel Hai_Final.Part 1
Jan 03, 2024
Abstract
A hospital readmission risk prediction tool for patients with diabetes based on electronic health record (EHR) data is needed.
With the improvement of people's living standards, diabetes has become a common chronic disease in our country. People with diabetes need to take good care of their bodies and constantly monitor and adjust their diet and living habits to ensure their health. In these aspects, the importance of memory is also highlighted.
Many studies have confirmed that diabetes and memory are inseparable. Diabetes affects the normal functioning of the brain, particularly memory, learning, and cognitive abilities. It has been found that patients with diabetes are more likely than ordinary people to suffer from memory decline, reduced learning ability, and weakened reaction ability.
However, we should not give up. Diabetes can be effectively controlled and its impact on our bodies through dietary and lifestyle changes.
First of all, dietary regulation is very important. People with diabetes need to limit their intake of sodium (salt), sugar, and fat to ensure a balanced food intake. A balanced dietary intake is beneficial to the body's sleep quality, mental health, hormonal balance, metabolism, and prevention of partial eclipse, all of which can help improve memory and cognitive abilities. Secondly, add some exercise. Not only does exercise help control blood sugar levels, but it can also help improve memory and cognitive abilities by strengthening muscles, reducing stress, and getting rid of bad moods.
Finally, proper treatment is required. Stable blood sugar levels also help improve the brain's learning and memory abilities.
Although people with diabetes may face various problems such as memory loss, this should not prevent people with diabetes from living a healthy and vibrant life. Monitoring your diet and lifestyle, doing some exercise, and maintaining an optimistic mood are very effective for health care. It can also better protect our body, memory, and cognition. We need to improve memory, and Cistanche deserticola can significantly improve memory because Cistanche deserticola is a traditional Chinese medicinal material with many unique effects, one of which is to improve memory. The efficacy of minced meat comes from the various active ingredients it contains, including acid, polysaccharides, flavonoids, etc. These ingredients can promote brain health in various ways.
Click know 10 ways to improve memory
The optimal modeling approach, however, is unclear. In 2,836,569 encounters with 36,641 diabetes patients, deep learning (DL) long short-term memory (LSTM) models predicting unplanned, all-cause, 30-day readmission were developed and compared to several traditional models. Models used EHR data defined by a Common Data Model.
The LSTM model Area Under the Receiver Operating Characteristic Curve (AUROC) was significantly greater than that of the next best traditional model [LSTM 0.79 vs Random Forest (RF) 0.72, p<0.0001]. Experiments showed that the performance of the LSTM models increased as the prior encounter number increased up to 30 encounters.
An LSTM model with 16 selected laboratory tests yielded equivalent performance to a model with all 981 laboratory tests. This new DL model may provide the basis for a more useful readmission risk prediction tool for diabetes patients.
Introduction
Hospital readmission is an undesirable, costly outcome for both patients and hospitals.1 Patients with diabetes are at higher risk of readmission within 30 days of hospital discharge (30-day readmission) than patients without diabetes.
2- 4 Of the nearly 9 million discharges of diabetes patients annually in the US, 5 almost 2 million are 30-day readmissions, corresponding to at least $20 billion in hospital costs.
6, 7 Identifying higher-risk patients with diabetes would enable the targeting of interventions to those in greatest need, optimizing the cost-benefit ratio.
We previously published the development and validation of the Diabetes Early Readmission Risk Indicator (DERRITM), a logistic regression (LR) model that predicts the risk of all-cause 30-day readmission among patients with diabetes.
8 The DERRITM was designed for use at the point of care based on user input of 10 factors. In split sample internal validation, performance was modest (Area Under the Receiver Operating Characteristic Curve, AUROC 0.69).
In external validation studies, the DERRITM AUROC was 0.63 and 0.80.9, 10 In addition to variable predictive performance, application of the DERRITM requires manual data collection and entry, which are major barriers to its use in clinical practice.
In other published work, we showed that adding variables to the DERRITM substantially improves predictive accuracy to an AUROC of 0.82.11 This expanded model (DERRIplus), however, is not feasible for use at the point of care and included employment status, which is not routinely documented in Electronic Health Records (EHRs).
Therefore, this model cannot be directly translated into an automated, EHR-integrated tool. There is an unmet need for a readmission risk prediction tool for patients with diabetes that is both accurate and easy to use.
Over the past few years, multiple machine learning (ML) models for predicting the 30-day readmission risk of diabetes patients have been published. Several traditional ML modeling approaches have been explored, including random forest (RF), k-nearest nearest neighbor, naïve Bayes, support vector machine (SVM), AdaBoost, and multilayer perceptron (MLP), with a wide range of performance (AUROC 0.53-0.99, accuracy 0.54-0.99).
12-22 Deep learning (DL) models have also been developed for predicting readmission risk of diabetes patients, also with variable performance (AUROC 0.61-0.97, accuracy 0.69-0.95), none of which exceeded that of the best traditional ML models.23-27 Two of these studies demonstrated a clear advantage of DL approaches over traditional ML models,23, 24, and two studies found marginal benefit with DL approaches.
25, 27 Comparisons of model performance across all these studies, however, are limited by the lack of standardized reporting of performance characteristics and variable approaches to testing.
Therefore, it remains unclear if DL models outperform traditional ML models at predicting readmission risk for patients with diabetes.
Interestingly, all these prior models were developed on the same dataset,28 except for the DERRITM and DERRIplus. This publicly available dataset contains hospital encounters with a diagnosis of diabetes and a length of stay between 1 and 14 days at one of 130 US hospitals between 1999 and 2008.
Only 3 International Classification of Diseases, Ninth Revision (ICD-9) diagnostic codes per encounter, and only 2 laboratory values (blood glucose and HbA1c) were recorded.
Lastly, there is no distinction made between planned and unplanned readmissions. Thus, even the best of these models may not perform as well in patients today. More current, generalizable models are needed.
Therefore, to address these gaps, the aims of the current study were as follows: 1) To develop DL models for the prediction of unplanned, all-cause 30-day readmission, 2) To compare the performance of the DL models to traditional ML models, 3)
To explore model performance across a range of prior EHR encounters from 1 to 100 being included in model development, and 4) To compare a DL model developed using a subset of laboratory tests selected by domain knowledge with a DL model developed using all available laboratory tests.
All models were developed and tested in a dataset of 2,836,569 encounters of 36,641 patients with diabetes using demographics, vital signs, diagnostic and procedure codes, medications, laboratory tests, and administrative data as defined by the National Patient-Centered Clinical Research Network (PCORnet) Common Data Model (CDM).29
Materials and Methods
Definition of the patient cohort
Inclusion criteria were patients with at least one discharge from any of the three Temple University Health System hospitals in Philadelphia, PA, between July 1st, 2010, and December 31st, 2020, and diabetes defined by at least one of the following: a diagnosis of diabetes (ICD-9: 249. xx or 250. xx or ICD-10: E08.xxx through E13.xxx); a Hemoglobin A1c (HbA1c) level ≥6.5%, or an order for a diabetes-specific medication.
Encounters were excluded for patients aged <18 years, discharged by transfer to another hospital, inpatient death, a diagnosis of gestational diabetes (ICD-9: 648.0x or ICD-10: O24.4x), a diagnosis of prediabetes (ICD-9: 790.29 or ICD-10: R73.03), or pregnancy (positive beta-human chorionic gonadotropin laboratory test within 90 days before or after the encounter).
Patients were sorted into one of 2 groups by readmission status: those who had at least one 30-day readmission and those who did not. Among the patients who had a readmission, one admission-readmission pair was randomly selected for analysis. Among the patients who did not have a readmission, one admission was selected randomly for analysis.
Definition of variables and data preprocessing
Tables were extracted from the CDM for each of the following domains: encounters, demographics, diagnoses, laboratory tests, medication orders, procedures, and vital signs. Because features of a given encounter existed in multiple tables, tables were merged by a unique identifier. Merging extracted tables resulted in a sample containing all records for a given encounter.
This resulted in substantial missingness. Thus, missingness was used as a separate feature. For continuous features, missing data were replaced with 0, while categorical features were replaced with a unique category.
A total of 23 features were used as input to the models: 14 were extracted from the CDM and 9 were aggregated. Extracted features were: 1) Encounter type (Inpatient, Emergency Department, Observation Stay, Ambulatory Visit, Other Ambulatory Visit, Telehealth and Other; 2) Discharge Status(Assisted Living Facility, Against Medical Advice, Expired, Home Health, Home/Self Care, Hospice, Nursing Home, Rehabilitation Facility, Skilled Nursing Facility; 3) Sex; 4) Hispanic; 5) Race (American Indian/Alaska Native, Asian, Black, Pacific Islander, White, other/no information); 6) Tobacco (Current user, never user, former user, passive exposure, other/no information); 7) age; 8)
Diagnosis Clinical Classification System (CCS) codes;29 9) Procedure CCS codes;29 10) Laboratory results; 11) Medication orders within 1 year before each encounter; 12) Diastolic blood pressure; 13) Systolic blood pressure; and 14) Body mass index (BMI).
Aggregated features were: 1) Elixhauser conditions: a binary feature indicating the presence or absence of each condition; 30 2) Duration of admission (length of stay in days); 3) several procedure codes before conversion to CCS code; 4) several diagnosis codes before conversion to CCS code; 5) number of days since the prior encounter regardless of encounter type; 6) several days since the prior inpatient, observation or emergency department encounter; 7) several days since the prior encounter of other (non-hospital) encounter types; 8) several inpatient, observation, and emergency department encounters before the current encounter; and 9) several other (non-hospital) encounters before the current encounter.
ICD-9 codes were converted to ICD-10 codes to unify the code format. ICD-10 codes and procedure codes were converted to CCS codes. Based on domain knowledge, medications relevant to diabetes were categorized as follows: diabetes medications by class, cholesterol, corticosteroids, renin-angiotensin system (RAAS) blood pressure agents, and non-RAS blood pressure agents. Other medications were ignored.
Features found not to be reliable, mostly missing, or correlated were removed. Outliers in features such as dates, results, height, weight, BMI, and blood pressure (systolic and diastolic) were removed by observing the data distributions, percentiles, and domain knowledge. Missing values were treated as another category that indicates that a parameter was not collected about the encounter.
The primary outcome for model prediction (�) was unplanned, all-cause inpatient readmission within 30 days of an inpatient encounter discharge as defined by the Centers for Medicare & Medicaid Services (CMS).31 Based on the CMS definition, only the first readmission within 30 days was analyzed.
For more information:1950477648nn@gmail.com
