Subjects
First, we screened 2,178 successive patients who were admitted to the University of Tokyo Hospital between 2009 and 2012. Among them, we selected patients with severe obesity who had a BMI of more than 35 kg/m2 and enrolled them in a comprehensive weight loss program (5,6). We excluded patients with severe cardiovascular disease, heart failure, infectious diseases, stroke, or peripheral artery disease, as well as patients with type 1 diabetes, pregnant women, patients with dementia, patients who had orthopedic diseases that could interfere with exercise (walking), perioperative patients, patients taking anti-obesity medications, patients who had undergone bariatric surgery, patients without pertinent data, patients who were transferred to another hospital immediately after discharge, patients who were readmitted, and those under 18 or over 80 years old. The remaining patients were observed for 3 years after discharge to assess subsequent weight gain and dropout. The comprehensive inpatient obesity treatment program targeted a 5% weight loss; Its details have been reported previously (7). It consisted of three main components, which include behavioral modification (goal setting and charting weight four times daily), diet, and exercise (patients with diabetes received appropriate anti-diabetic therapy together with this weight loss program, and the treatment of diabetes was decided by each attending physician).
Behavioral modification (goal setting and charting weight four times daily)
At admission, all patients were given the goal of achieving 5% weight loss while in the hospital. They also weighed themselves four times daily (immediately after waking, taking breakfast and dinner, and before going to bed) and recorded the data on a weekly graph. A daily record of weight fluctuations reveals irregular intake of food and fluid that reflects dysfunctional eating patterns and other behavioral abnormalities and assists in achieving weight loss (5,8,9). Patients were recommended to continue recording their weight after discharge. Patients were weighed by using AD-6107NWTM scales (A and D Co. Ltd., Tokyo, Japan).
Diet
A balanced low-calorie diet (20–24 kcal/day/kg of ideal body weight) was provided to the patients in the hospital, consisting of 50%-60% carbohydrate, 20% protein, and 20%-30% fat. Hospital dietitians used food samples and a food exchange table (10) to educate patients about nutritional guidelines. The dietician initially gave each patient information for 1 hour, with subsequent 30-minutes sessions being held at least twice a week until discharge.
Exercise
All patients were given a pedometer and were instructed to walk more than 10,000 steps/daily (5–7 km) for approximately 1.5 hours. The exercise program was tailored to accommodate health problems (e.g., morbid obesity, hypoglycemia, joint pain, or diabetic retinopathy) and specific needs (e.g., exercise by walking or training on a bicycle ergometer). The target pulse rate and schedule for each exercise session were set.
Outcome measures
The patients attended our hospital outpatient department every 2 months after discharge to continue their weight loss program and for treatment of other diseases (diabetes, dyslipidemia, and hypertension).
The body weight of the patients was measured at each visit. The objective of this study was to assess the dropout from the weight loss program after discharge. Dropout from the program was defined as missing outpatient appointments. (If the patient presented again within 6 months of the specified appointment, this was not considered as dropout.)
Patients were defined as having diabetes if their medical records listed a diagnosis of type 2 diabetes, and they were using an oral hypoglycemic agent or insulin. If a 75 g oral glucose tolerance test was performed, a diagnosis of diabetes, impaired glucose tolerance, or impaired fasting glucose was made according to the American Diabetes Association criteria (11). Antidepressant medications were classified as selective serotonin reuptake inhibitors, serotonin and norepinephrine reuptake inhibitors, tricyclic antidepressants, tetracyclic antidepressants, serotonin receptor antagonists and reuptake inhibitors, monoamine oxidase inhibitors, and noradrenergic and specific serotonergic antidepressants. All demographic and clinical data were collected from secure electric medical records. Nurses or physicians confirmed the accuracy of the bodyweight measurements of each subject
Feature engineering
We used binary variables and continuous variables as prediction features. The binary variables included gender, use of oral antidiabetic drugs, use of insulin, use of Glucagon-like peptide-1 Receptor analogs (GLP1-Ras), diabetes, hypertension, psychiatric disease, depression, insomnia, and antipsychotic drug use. The continuous variable included sequential body weight data from baseline to day 14, discharge body weight, age, waist circumference, systolic blood pressure, and HbA1C.
Data preparation
Overall, 102 patients (82 non-dropout and 20 dropout) were enrolled in this study. Due to the original imbalance sample between the dropout and non-dropout population, we used the Adaptive Synthetic Sampling (ADASYN) method to deal with the imbalanced data (12). In the balanced sample generated by the ADASYN method, there were 79 dropout and 82 non-dropout events. We randomized 85% of this balanced sample to a training cohort (for model training) and the rest of 15% as a validation cohort (for hyperparameter optimization). Then, we applied the deep learning (DL) algorithm generated from the balanced sample back to the original imbalanced cohort for individual prediction. The logistic regression (LR) model was also derived from the balanced sample and applied back to the original 102 people as individual risk prediction.
Model building process of deep learning
A machine learning model to predict the dropout rate was developed with deep neural networks. The variables derived from feature engineering (Table 1) were used as predictors and were set as input layers. For better performance of deep learning, we performed batch normalization (to mean 0 or variance 1) for the selected variables (features) and set this as the input layer.
Table 1
Baseline characteristics between dropout and non-dropout at the end of the first year.
| | | | |
Variable | Overall | Non-dropout(1y) | dropout(1y) | p-value |
N | 102 | 82 | 20 | |
Age (years old) | 49.3 ± 14.1 | 50.3 ± 13.8 | 45 ± 15.1 | 0.13 |
Male (%) | 43(42) | 35(43) | 8(40) | 0.83 |
BMI | 42 ± 9 | 41 ± 9 | 41 ± 5 | 0.9 |
Wait circumstance (cm) | 123 ± 13.2 | 122 ± 13.1 | 126.9 ± 12.9 | 0.13 |
Systolic blood pressure (mmHg) | 131.3 ± 17 | 131.8 ± 16.5 | 129.2 ± 19.4 | 0.54 |
HbA1c (%) | 7.6 ± 2.5 | 7.7 ± 2.4 | 6.8 ± 2.5 | 0.13 |
Type 2 diabetes (%) | 78(76.47) | 67(81.71) | 11(55) | 0.01 |
Hypertension (%) | 68(66.67) | 59(71.95) | 10(50) | 0.06 |
Psychosis (%) | 41(40.2) | 33(40.24) | 8(40) | 0.98 |
Depression (%) | 14(13.73) | 11(13.41) | 3(15) | 0.85 |
Insomnia (%) | 24(23.53) | 21(25.61) | 3(15) | 0.32 |
Other mental disease (%) | 25(24.51) | 19(23.17) | 6(30) | 0.52 |
Oral hypoglycemic agent (%) | 51(50) | 45(54.88) | 6(30) | 0.046 |
Insulin therapy (%) | 23(22.55) | 20(24.39) | 3(15) | 0.37 |
Sulfonylurea | 5(4.9) | 5(6.1) | 0(0) | 0.26 |
GLP-1 analogs | 14(13.73) | 13(15.85) | 1(5) | 0.21 |
The entire structure of the deep neural network was designed as follows: 25 input layers → 25 middle hidden layers → 25 middle hidden layers → 20 middle hidden layers → 16 middle hidden layers → one-dimensional output layer. The binary outcome of the 1-year dropout or non-dropout was set as the output layer. To avoid overfitting during the model training of deep learning, we used an L1 regularization method (13), as well as adding a dropout layer between the hidden layers. The dropout rate was set at 0.2. As activation functions, we employed scaled exponential linear units in the middle layer and hard sigmoid units in the output layer.
Statistics
The independent t test and χ2 test were used to compare continuous and categorical variables between dropout and non-dropout subgroups. To compare the model performance between DL and LR, we calculated the area under the receiver operating characteristic curve (ROC) based on the 1-year dropout event in the original cohort (N = 102). To see the extension of the prediction ability after 1 year, we plotted the 3-year survival curve using the Kaplan-Meir method according to DL and LR classification. We used a TensorFlow version 1.10.0. for model building. Logistic regression analysis was performed with Scikit-learn version 0.20.3. Machine learning was done with Python version 3.6.5 (Python Software Foundation, Hampton, NH). Statistical analyses were performed using the SAS software (Version 9.4; SAS Institute, Inc., Cary, NC, USA).