Predicting suicide attempts among people with diabetes using a large multicenter electronic health records dataset

Abstract

Objective

People with diabetes have a higher risk of suicidal behaviors than the general population. However, few studies have focused on understanding this relationship. We investigated risk factors and predicted suicide attempts in people with diabetes using Least Absolute Shrinkage and Selection Operator (LASSO) regression.

Method

Data was retrieved from Cerner Real-World Data and included over 3 million diabetes patients in the study. LASSO regression was applied to identify associated factors. Gender, diabetes-type, and depression-specific LASSO regression models were analyzed.

Results

There were 7764 subjects diagnosed with suicide attempts with an average age of 45. Risk factors for suicide attempts in diabetes patients were American Indian or Alaska Native race (β = 0.637), receiving atypical antipsychotic agents (β = 0.704), benzodiazepines (β = 0.784), or antihistamines (β = 0.528). Amyotrophy was negatively associated with suicide attempts in males (β = 2.025); in contrast, amyotrophy significantly increased the risk in females (β = 3.339). Using a MAOI was negatively related to suicide attempts in T1DM patients (β = 7.304). Age less than 20 was positively associated with suicide attempts in depressed (β = 2.093) and non-depressed patients (β = 1.497). The LASSO model achieved a 94.4% AUC and 87.4% F1 score.

Conclusions

To our knowledge, this is the first study to use LASSO regression to identify risk factors for suicide attempts in patients with diabetes. The shrinkage technique successfully reduced the number of variables in the model to improve the fit. Further research is needed to determine cause-and-effect relationships. The results may help providers to identify high-risk groups for suicide attempt among diabetic patients.

Keywords

diabetes least absolute shrinkage and selection operator risk factors suicide attempts

Introduction

Diabetes mellitus (DM) is a chronic non-communicable disease characterized by high blood sugar. People with diabetes have higher rates of depression, anxiety, and severe psychological distress than non-diabetic people.^1-4 In 2020, the National Survey on Drug Use and Health reported that 1.2 million Americans aged 18 and older, or 0.5% of the US population, attempted suicide in the past year.⁵ Research has shown that people with type 1 diabetes mellitus (T1DM) are three to four times more likely to attempt suicide than the general population.⁵ Similarly, people with newly diagnosed type 2 diabetes mellitus (T2DM) have a rate of suicide attempts two times greater than the general population.⁶ Suicide attempts are significantly associated with the diagnosis of depressive disorders, which is prevalent in people with diabetes.^7,8 Studies have found that 38.2% of adults with T2DM were diagnosed with major depressive disorder, and the prevalence of suicide attempts among these adults was significantly higher (21.8%) than that of the general adult population (0.5%).⁸

While suicide attempts significantly impact people with diabetes, only a few studies have focused on understanding the relationship between suicide attempts and diabetes.⁹ Previous studies applied logistic regression analysis to identify factors that increased the risk of attempted suicide among people with diabetes, such as females,^5,10 younger age,⁹ normal body mass index (BMI),⁹ depression,^5,10,11 history of childhood abuse,⁵ history of alcohol abuse,⁵ and higher insulin in cerebrospinal fluid.¹²

It is widely acknowledged that the comorbidity of depression with diabetes is considered a higher risk factor for suicide attempts (HR = 5.64; 95% CI 4.70–6.77).^13-18 However, there are several limitations to using logistic regression analysis. First, the number of observations should be more than the number of features. Second, logistic regression may not work well with a complex relationship when many variables are considered in the model, such as medical data, because a high number of variables in the logistic regression can lead to overfitting.¹⁹ LASSO (Least Absolute Shrinkage and Selection Operator) is a regularization technique used to reduce overfitting in a logistic regression model. It works by adding a penalty term (λ) to the log-likelihood function, which shrinks the estimates of the coefficients. This process allows LASSO to act as a feature selection method, effectively setting coefficients that contribute most to the error to zero.^20-22

The aim of this study was to identify factors associated with attempted suicide in individuals with diabetes. To achieve this goal, we first conducted a literature review to identify potential risk factors for inclusion as variables in our study. We then used multivariate and LASSO regression analyses to determine which factors were statistically significant in predicting suicide attempts among people with diabetes, using data from the Cerner Real-World Data™ (CRWD). Our long-term objective is to develop a clinical decision support system for healthcare providers to identify patients with diabetes who are at high risk of attempting suicide. This system would aid in providing suicide prevention measures during outpatient clinic visits.

Method

Risk factor identification via literature review

We conducted a literature review to identify potential risk factors for suicide attempts in people with diabetes and created a list of related data elements. The first objective of the literature review was to identify risk factors for suicide attempts among people with diabetes (Figure 1). We searched articles from three databases, including PubMed, Scopus, and PsycINFO, using MeSH terms: “suicide, attempted,” “diabetes mellitus, type 1”, “diabetes mellitus, type 2,” and “risk factors.” The results were limited to 31 December 2021. There were few studies focusing on identifying risk factors for suicide attempts and diabetes. Thus, we expanded the results by using “suicide” instead of “suicide, attempted” based on the MeSH hierarchy. Articles were included if they met the following criteria: (1) observational studies, including cross-sectional, case-control, and cohort studies; (2) suicidal behaviors diagnosed by standardized suicide behaviors assessment tools; (3) reporting the association between T1DM or T2DM and suicide; (4) studies reporting specific risk factors of comorbid suicide with T1DM or T2DM; (5) the subjects that were diagnosed with T1DM or T2DM at the baseline; (6) studies published in English. We excluded articles if they met the following criteria: (1) studies reporting risk factors for neither suicide nor diabetes (T1DM and T2DM); (2) unavailable full texts; (3) reviews, case reports, case series, letters to the editor, and abstract publications; (4) gestational diabetes. We excluded gestational diabetes because the pathophysiology of gestational diabetes is different from T1DM and T2DM.²³ Also, the physiology of pregnant women changes throughout the pregnancy.²⁴

Figure 1.

Flow diagram of literature review to identify risk factors of suicidal behavior among people with diabetes.

The second objective of the literature review was to identify risk factors for depression among people with diabetes (Figure 2). We found that previous studies focusing on the relationship between suicide attempts and diabetes were limited, and depression is a major risk factor for suicide attempts among people with diabetes. Therefore, we considered the possible risk factors, not only the relationship between suicide attempts and diabetes but also the relationship between depression and diabetes. We searched articles from three databases, including PubMed, Scopus, and PsycINFO, using MeSH terms: “depression,” “diabetes mellitus, type 1”, “diabetes mellitus, type 2,” and “risk factors.” The results were limited to 31 December 2021. Last, we combined the results from both literature reviews and created a list of candidate risk factors.

Figure 2.

Flow diagram of literature review to identify risk factors of depression among people with diabetes.

Data collection

We extracted data elements from Cerner Real-World Data (CRWD) from 2010 to 2020 using the International Classification of Diseases, Ninth/Tenth Revision, Clinical Modification (ICD-9/10-CM). CRWD is a large data source consisting of approximately 100 million de-identified patients and encompasses 1.5 billion HIPAA-compliant patient encounters from 117 hospitals’ electronic health records (EHR) across the United States. CRWD contains longitudinal data, including demographics, encounters, clinical events, conditions, immunization, labs, measurements, medications, order lists, procedures, and results. The diagnoses of patients use coding systems, including ICD-9/10-CM and Systemized Nomenclature of Medicine - Clinical Terms (SNOMED).²⁵ A quantitative analysis was performed to determine the prevalence and risk factors of suicide attempts among diabetes patients. We limited data from 2010 to 2020 to control the data quality because we found that data before 2010 had incomplete information and timestamp errors. To develop our dataset, we extracted CRWD through HealtheDataLab™, a data science ecosystem built and deployed on Amazon Web Services.²⁶ HealtheDataLab™ uses Jupyter notebook.²⁷ as a web-based interactive interface for analysis and offers open-source tools, such as Structured Query Language (SQL), Python, and R in the Apache Spark environment.²⁸

Statistical analysis

We determined the minimum number of study participants needed to achieve sufficient statistical power. Our calculations were based on an estimated population prevalence of 7.6%, a prevalence of 0.2% in the study group, a significance level (Alpha) of 0.05, a type II error rate (Beta) of 0.05, and a desired power of 0.95. The results of our analysis indicate that a minimum of 64 participants is required for the study. In order to identify potential risk factors, our dataset was analyzed using R programming language version 4.0.2 on the HealtheDataLab™ platform. We employed Chi-square tests with 95% confidence intervals to assess the association between different variables and suicide attempts among patients with diabetes. Variables with more than 50% missing values were removed from the analysis.²⁹ Missing values of less than 50% were imputed using Multivariate Imputation by Chained Equations (MICE) method.³⁰ The dataset was divided into two sets, with 90% used for training and 10% for testing. Random Over-Sampling Examples (ROSE), a bootstrap-based technique, was used to address the problem of class imbalance in the training dataset, i.e. when there is an unequal distribution of a class of suicide attempts and a class of non-suicide attempts. Using an imbalanced dataset may cause a loss of necessary information from the minority class to build the model, resulting in biased prediction and misleading accuracy. ROSE over-samples the minority class (suicide attempts) by randomly selecting and duplicating instances until the class distribution is balanced. This process creates a new training dataset with equal numbers of instances for both classes, allowing the classifier to learn both classes equally.³¹

To improve the performance and determine the most suitable model for our study, we developed two predictive models. First was a logistic regression model fitted using the “glmnet” package and the “glm” function. This model was used to determine risk factors by analyzing the association between variables and suicide attempts among patients with diabetes and determining the adjusted odds ratios (aOR) with 95% confidence intervals (CI). The model was then tested for prediction using the test data set.

The second model was a LASSO regression model, which was fitted using the “cv.glmnet” function. This model used feature selection to minimize the overfitting problem caused by a high number of variables. LASSO regularization was applied by adding a penalty term (λ) to the log-likelihood function and setting the coefficients that contributed most to the error to zero. Ten-fold cross-validation was used to select the largest λ within one standard error of the minimum binomial deviance. The best cross-validated λ was 0.00079, used in the final model to determine risk factors with coefficients.

To interpret the results from the LASSO regression model, the magnitude of the coefficients ( $β$ ) was used to determine the strength of the association between the feature and the target variable. Features with larger magnitude coefficients are considered to be more important. We also considered the direction of the coefficient when we evaluated the association between the feature and the target variable. A positive coefficient indicates a positive association, while a negative coefficient indicates a negative association. If a coefficient is close to zero, this suggests that the corresponding feature is not strongly associated with the response variable. Features with coefficients close to zero can be considered as having little or no predictive power.²¹ To evaluate the performance of the models, we tested the LASSO model for prediction with the testing dataset. Confusion matrix, F1 score, Area Under the Curve (AUC), accuracy, precision, and recall were calculated.

In order to further understand the potential risk factors for suicide attempts among patients with diabetes, we conducted subgroup analyses specific to gender and diabetes type. Some studies have suggested that women may have a higher risk of suicide attempts than men due to a higher prevalence of certain mental disorders such as anxiety, depression, and eating disorders.³² However, other studies have presented contradictory findings.³² Similarly, there is conflicting evidence on whether patients with T1DM or T2DM have a higher risk of suicide.³³ To address these inconsistencies, we utilized LASSO regression analysis to analyze the differences between gender, diabetes type, and depression subgroups.

Results

Patients’ characteristics

After querying data using ICD-9/10 and text searching, we identified 52,217,517 unique patients in the CRWD. A total of 3,266,856 individual diabetes patients were included in our dataset from 01-01-2010 to 31-12-2020. Of 3,266,856 patients diagnosed with diabetes, 7764 (0.2%) were diagnosed with suicide attempts. For the non-diabetic group, 80,035 of 42,324,121 patients (0.2%) were diagnosed with suicide attempts. The average age of diabetes patients with suicide attempts was 45 years (SD = 16, 8-88), and 47% were aged 41–60. Approximately half of them were female (57%), never married (52%), and. BMI ≥30 kg/m² (54%). Most of them were white (74%) and non-Hispanic (74%). The prevalence of diabetes patients with suicide attempts who had a family history of psychiatric illness was 9%. Less than 5% of them had a death of family members or divorce and a history of childhood abuse.

LASSO regression analysis

The performances of the multivariate regression model were 94.4% AUC, 92.7% accuracy, and 3.3% F1 score. The performances of the LASSO regression model were 94.4% AUC, 87.5% accuracy, and 87.4% F1 score. The results from the multivariate and LASSO regression models are shown in Table 1. We divided risk factors for suicide attempts among people with diabetes from the LASSO model into five groups: (1) Patients’ characteristics included age ≤20, age 21-40, male, never married, separated, White, Asian, American Indian or Alaska Native, Hispanic or Latino, BMI 18.5-24.9 kg/m², family history of psychiatric illness, death of family members or divorce, history of childhood abuse, and T1DM; (2) Diabetes complications included amyotrophy, hyperglycemia, ketoacidosis, neuropathy, diabetes with other coma, retinopathy, and unspecified complication; (3) Comorbidities included anxiety, cognitive impairment, depression, eating disorder, hypertension, hypothyroidism, ischemic heart disease, and myocardial infarction; (4) Medication profiles included insulin, selective serotonin reuptake inhibitors (SSRI), serotonin and norepinephrine reuptake inhibitors (SNRI), monoamine oxidase inhibitors (MAOI), atypical agents, azaperone, benzodiazepines and antihistamines; (5) Laboratory profiles included average A1C <5.7%, average A1C ≥9%, average random glucose <150 mg/dl, average low-density lipoprotein (LDL) <100 mg/dl and average triglycerides (TG) 150-199 mg/dl.

Table 1.

Risk factors for suicide attempts in diabetes patients comparing between models.

	Multivariate		LASSO
	Adjusted OR (95% CI)	p-value	Coefficient
Patient characteristics
Age (years)
≤20	22.04 (21.56-22.53)	<0.001	1.544
21–40	12.44 (12.27-12.62)	<0.001	0.999
41–60	4.41 (4.37-4.46)	<0.001
>60	NA	NA	−1.457
Gender
Male	0.49 (0.44-0.55)	<0.001	0.069
Female	0.76 (0.68-0.85)	0.000	−0.340
Marital status
Married	1.32 (1.3-1.35)	<0.001	−0.559
Never Married	0.74 (0.73-0.75)	<0.001	0.047
Divorced	1.37 (1.35-1.39)	<0.001
Separated	0.57 (0.54-0.6)	<0.001	0.206
Widowed	1.67 (1.63-1.72)	<0.001	−0.245
Race
White	1.86 (1.81-1.91)	<0.001	0.076
Black or African American	1.01 (0.98-1.04)	0.504	−0.149
Asian	0.79 (0.78-0.8)	<0.001	0.015
American Indian or Alaska Native	0.93 (0.88-0.99)	0.021	0.637
Native Hawaiian or Other Pacific Islander	0.9 (0.89-0.92)	<0.001
Ethnicities
Hispanic or Latino	1.09 (1.07-1.11)	<0.001	0.064
BMI (kg/m2)
<18.5	0.67 (0.65-0.7)	<0.001	−0.541
18.5–24.9	1.28 (1.26-1.3)	<0.001	0.011
25–29.9	1.26 (1.25-1.27)	<0.001
≥30	NA	NA	−0.221
Family history of psychiatric illness	4.36 (4.11-4.62)	<0.001	0.947
Death of family members or divorce	3.12 (2.99-3.25)	<0.001	1.021
History of childhood abuse	0.82 (0.81-0.83)	<0.001	1.252
Family history of diabetes	2.81 (2.72-2.91)	<0.001	-0.149
Type 1 diabetes	1.18 (1.15-1.2)	<0.001	0.154
Diabetes duration <5 years	0.97 (0.96-0.99)	0.000	-0.005
Diabetes complications
Amyotrophy	5.1 (4.09-6.36)	<0.001	1.131
Arthropathy	0.91 (0.83-0.99)	0.027
Foot ulcer	0.29 (0.28-0.31)	<0.001	−1.134
Hyperglycemia	1.09 (1.08-1.11)	<0.001	0.041
Hyperosmolarity	0.86 (0.81-0.91)	0.000	−0.097
Hypoglycemia	0.87 (0.85-0.89)	<0.001	−0.118
Ketoacidosis	1.43 (1.4-1.47)	<0.001	0.288
Nephropathy	0.86 (0.84-0.87)	<0.001	−0.160
Neuropathy	2.07 (2.03-2.1)	<0.001	0.622
Other coma	8.96 (7.75-10.35)	<0.001	1.829
Peripheral circulatory disorders	0.62 (0.61-0.64)	<0.001	−0.462
Retinopathy	1.15 (1.11-1.18)	<0.001	0.024
Skin complications	0.46 (0.43-0.48)	<0.001	−0.713
Unspecified	3.87 (3.83-3.9)	<0.001	1.305
Comorbidities
Anxiety	2.77 (2.75-2.8)	<0.001	1.011
Cerebral infarction	1 (0.99-1.02)	0.738
Cognitive impairment	1.3 (1.28-1.33)	<0.001	0.236
Depression	9.03 (8.95-9.11)	<0.001	2.185
Eating disorder	0.48 (0.47-0.5)	<0.001	0.667
Erectile dysfunction	0.34 (0.32-0.36)	<0.001	−0.674
Hypertension	1.3 (1.29-1.31)	<0.001	0.228
Hypothyroidism	1.08 (1.06-1.09)	<0.001	0.053
Ischemic heart disease	1.1 (1.09-1.11)	<0.001	0.075
Myocardial infarction	1.06 (1.05-1.08)	0.000	0.047
Obesity	0.8 (0.79-0.81)	<0.001	−0.196
Medication profiles
Diabetes medications
Insulin	1.09 (1.08-1.11)	<0.001	0.072
Biguanides	0.70 (0.69-0.71)	<0.001	−0.183
Sulfonylureas	0.89 (0.88-0.91)	<0.001	−0.009
Thiazolidinedione	0.61 (0.59-0.62)	<0.001	−0.393
DPP-4 inhibitor	0.83 (0.82-0.85)	<0.001	−0.098
SGLT2 inhibitor	0.6 (0.59-0.61)	<0.001	−0.419
GLP-1 receptor agonist	0.55 (0.54-0.56)	<0.001	−0.495
Antidiabetic combinations	0.68 (0.66-0.69)	<0.001	−0.315
Antidepressants
SSRI	1.43 (1.41-1.44)	<0.001	0.301
SNRI	1.85 (1.83-1.88)	<0.001	0.533
Tricyclic	1.02 (1-1.04)	0.015
MOI	4.78 (4.1-5.58)	<0.001	1.211
Atypical agents	2.2 (2.17-2.23)	<0.001	0.704
Anti-anxiety medications
Azaperone	1.14 (1.12-1.17)	<0.001	0.048
Benzodiazepines	2.26 (2.24-2.28)	<0.001	0.784
Antihistamines	1.86 (1.82-1.89)	<0.001	0.528
Anticonvulsants	0.94 (0.92-0.96)	0.000	−0.092
Laboratory profiles
Average A1C (%)
<5.7	1.03 (1.01-1.05)	0.001	0.182
5.7-6.4	0.85 (0.83-0.86)	<0.001
6.5-8.9	0.83 (0.82-0.85)	<0.001	−0.020
≥9	NA	NA	0.134
Average random glucose (mg/dl)
<150	1.78 (1.68-1.88)	<0.001	0.081
151-200	1.32 (1.24-1.39)	<0.001	−0.191
201-300	1.59 (1.5-1.69)	<0.001
301-400	1.15 (1.08-1.22)	0.000	−0.276
≥401	NA	NA	−0.369
Average cholesterol (mg/dl)
<200	0.98 (0.96-1)	0.012
≥200	NA	NA
Average HDL (mg/dl)
Female ≥ 50, Male ≥ 40	0.98 (0.97-0.99)	0.000	−0.008
Female < 50, Male < 40	NA	NA	0.000
Average LDL (mg/dl)
<100	1.44 (1.36-1.52)	<0.001	0.148
100-129	1.23 (1.16-1.3)	0.000
130-159	1.07 (1.01-1.13)	0.020	−0.088
160-189	1.07 (1-1.14)	0.043	−0.052
≥190	NA	NA	−0.070
Average TG (mg/dl)
<150	1.29 (1.25-1.33)	<0.001	−0.058
150-199	1.47 (1.42-1.52)	<0.001	0.048
200-499	1.39 (1.35-1.44)	<0.001
≥500	NA	NA	−0.244

Gender-specific analysis

Male-specific LASSO regression model had 89.5% accuracy and 93.3% AUC. Depression had the highest coefficient (

β

= 2.304), and amyotrophy had the lowest coefficient (

β

= −2.025). Female-specific LASSO regression model had 86.5% accuracy and 95% AUC. Amyotrophy had the highest coefficient (

β

= 3.339), and age >60 years had the lowest coefficient (

β

= −1.753). We found some factors were associated with suicide attempts in males with diabetes, such as being divorced, arthropathy, average A1C 6.5-8.9%, average cholesterol <200 mg/dl, and average TG 150-199 mg/dl. In contrast, those factors were not associated with suicide attempts in females with diabetes (Table 2).

Table 2.

Coefficients of gender-, type of diabetes-, and depression-specific LASSO regression analysis.

	Male	Female	T1DM	T2DM	Depression	Non-depression
	(Accuracy = 0.895, AUC = 0.933)	(Accuracy = 0.865, AUC = 0.950)	(Accuracy = 0.843, AUC = 0.895)	(Accuracy = 0.879, AUC = 0.944)	(Accuracy = 0.440, AUC = 0.861)	(Accuracy = 0.960, AUC = 0.893)
Patient characteristics
Age (years)
≤20	0.984	1.845	1.167	1.647	2.093	1.497
21-40	1.035	0.949		1.021	0.945	1.047
41-60			−0.452
>60	−1.171	−1.753	−2.761	−1.453	−1.495	−1.415
Gender
Male	X	X	0.211	0.039	0.187	.
Female	X	X	.	−0.383	-0.361	−0.372
Marital status
Married	−0.677	−0.412	−0.207	−0.561	−0.572	−0.496
Never Married	0.041	0.111	.	0.066	.	0.120
Divorced	0.079	.	0.938	.	.	.
Separated	0.296	0.180	.	0.230	.	0.005
Widowed	−0.023	−0.168	0.372	−0.245	−0.171	0.467
Race
White	0.067	0.109	0.271	0.090	.	0.230
Black or African American	−0.140	−0.112	.	−0.148	−0.195	.
Asian	.	0.168	−0.732	0.035	0.378	.
American Indian or Alaska Native	0.460	0.802	.	0.638	0.686	0.672
Native Hawaiian or Other Pacific	.	0.112	−0.438	.	0.268	-0.081
Islander
Ethnicities
Hispanic or Latino	0.079	0.043	−0.245	0.066	0.168	0.011
BMI (kg/m2)
<18.5	−0.935	−0.132	−2.724	−0.442	−0.300	−0.642
18.5–24.9	.	0.014	.	0.009	0.048	.
25–29.9	.	.	0.097	.	.	0.024
≥30	−0.235	−0.208	−1.312	−0.200	−0.213	−0.220
Family history of psychiatric illness	1.107	0.838	0.442	0.957	0.873	1.276
Death of family members or divorce	0.960	0.981	1.284	0.995	0.922	1.527
History of childhood abuse	1.066	1.266	0.191	1.288	1.126	1.832
Family history of diabetes	−0.061	−0.166	−0.167	−0.152	−0.111	−0.155
Types of Diabetes
Type 1	0.244	0.101	X	X	0.794	.
Type 2	.	0.000	X	X	0.000	.
Diabetes duration (years)
<5	.	−0.014	−0.138	−0.002	0.041	−0.157
≥5	.	0.000	0.000	.	0.000	0.000
Diabetes complications
Amyotrophy	−2.025	3.339	−3.039	1.250	−3.660	2.712
Arthropathy	−0.028	.	.	.	.	−0.332
Foot ulcer	−1.045	−1.531	.	−1.252	−1.284	−1.091
Hyperglycemia	.	0.047	1.564	0.024	−0.066	0.105
Hyperosmolarity	.	−0.346	.	−0.113	−0.391	.
Hypoglycemia	.	−0.239	−0.178	−0.075	−0.108	−0.232
Ketoacidosis	0.259	0.321	2.264	0.126	.	0.458
Nephropathy	−0.282	−0.048	−0.653	−0.167	−0.238	−0.060
Neuropathy	0.479	0.715	2.538	0.591	0.506	0.695
Other coma	0.776	2.545	4.317	1.763	1.220	2.305
Peripheral circulatory disorders	−0.557	−0.393	−0.527	−0.470	−0.717	−0.050
Retinopathy	0.160	−0.037	1.883	.	−0.223	0.363
Skin complications	−1.233	−0.110	2.331	−0.778	−0.909	−0.516
Unspecified	1.194	1.404	2.001	1.296	1.302	1.327
Comorbidity
Anxiety	0.858	1.154	0.976	1.011	0.701	1.282
Cerebral infarction	.	.	−2.317	.	−0.061	.
Cognitive impairment	0.170	0.270	−0.903	0.254	0.170	0.217
Depression	2.304	2.069	3.393	2.167	X	X
Eating disorder	1.208	0.380	0.247	0.566	0.551	0.771
Erectile dysfunction	−0.733	.	−1.056	−0.694	−0.809	−0.704
Hypertension	0.293	0.164	0.164	0.228	0.167	0.224
Hypothyroidism	0.042	0.088	.	0.050	0.000	0.068
Ischemic heart disease	.	0.130	0.275	0.070	0.112	−0.091
Myocardial infarction	.	0.053	.	0.039	0.011	−0.056
Obesity	−0.147	−0.238	.	−0.207	−0.268	−0.183
Medication profiles
Diabetes medications
Insulin	0.006	0.091	−0.229	0.103	0.087	0.071
Biguanides	−0.224	−0.176	0.765	−0.197	−0.155	−0.201
Sulfonylureas	−0.034	−0.089	−0.360	−0.007	−0.146	0.042
Thiazolidinedione	−0.487	−0.353	−4.952	−0.389	−0.201	−0.645
DPP-4 inhibitor	−0.125	−0.071	−1.874	−0.109	.	−0.228
SGLT2 inhibitor	−0.387	−0.499	−4.845	−0.403	−0.261	−0.655
GLP-1 receptor agonist	−0.441	−0.605	−1.292	−0.498	−0.549	−0.439
Antidiabetic combinations	−0.174	−0.544	−7.150	−0.308	−0.202	−0.437
Antidepressants
SSRI	0.313	0.287	0.952	0.285	.	0.670
SNRI	0.557	0.516	0.658	0.532	0.292	0.976
Tricyclic	.	.	−0.828	.	−0.069	.
MOI	0.895	1.272	−7.304	1.242	1.042	1.272
Atypical agents	0.668	0.717	0.658	0.701	0.475	0.912
Anti-anxiety medications
Azaperone	.	0.082	−0.027	0.041	0.068	0.226
Benzodiazepines	0.832	0.737	1.441	0.772	0.644	0.886
Antihistamines	0.428	0.587	.	0.541	0.566	0.541
Anticonvulsants	.	−0.155	−0.868	−0.080	−0.045	−0.057
Laboratory profiles
Average A1C
<5.7	0.108	0.222	.	0.188	0.224	0.100
5.7-6.4	.	.	0.940	.	.	.
6.5-8.9	−0.080	.	−0.496	−0.008	−0.001	−0.082
≥9	0.001	0.209	.	0.128	0.184	0.032
Average random glucose
<150	0.001	0.142	−0.339	0.117	.	0.240
151–200	−0.234	−0.139	.	−0.161	−0.020	−0.269
201–300	.	.	0.428	.	0.028	.
301–400	−0.237	−0.203	−0.264	−0.276	−0.379	−0.088
≥401	−0.072	−0.750	0.454	−0.492	−0.788	−0.129
Average cholesterol
<200	0.083	.	−0.330	.	0.028	.
≥200	.	.	0.000	.	0.000	.
Average HDL
Female≥50, Male≥40 (Desirable)	.	−0.008	−0.005	−0.004	.	−0.004
Female<50, Male<40 (At risk)	.	0.001	0.000	0.000	.	.
Average LDL
<100	0.144	0.098	0.229	0.138	0.112	0.151
100-129	.	.	−0.703	.	.	.
130-159	−0.138	−0.070	.	−0.101	−0.155	.
160-189	−0.027	−0.049	0.709	−0.118	.	.
≥190	.	−0.225	−0.545	−0.111	−0.054	−0.002
Average TG
<150	.	0.039	−0.691	−0.054	.	−0.132
150-199	0.190	.	0.190	0.040	0.007	0.086
200-499	−0.010	0.154	.	.	.	.
≥500	.	−0.265	.	−0.246	−0.058	−0.343

Diabetes type-specific analysis

T2DM-specific LASSO regression model had 87.9% accuracy and 94.4% AUC. Depression had the highest coefficient ( $β$ = 2.167), and age >60 years had the lowest coefficient ( $β$ = −1.453). T1DM-specific LASSO regression model had 84.3% accuracy and 89.5% AUC. Diabetes with other coma complications had the highest coefficient ( $β$ = 4.317), and the use of MAOI had the lowest coefficient ( $β$ = −7.304). We found some factors were associated with suicide attempts in T2DM but were not associated with suicide attempts in T1DM, such as female, age 20-40 years, Black or African American, American Indian or Alaska Native, never married, separated, BMI 18.5-24.9 kg/m², hyperosmolarity, foot ulcer, obesity, hypothyroidism, myocardial infarction, antihistamines, A1C <5.7%, A1C ≥9%, average random glucose 151-200 mg/dl, LDL 130-159 mg/dl, and average TG ≥500 mg/dl (Table 2).

Depression-specific analysis

The LASSO regression model for depressed diabetes patients had 44.0% accuracy and 86.1% AUC. Age ≤20 years had the highest coefficient ( $β$ = 2.093), and amyotrophy had the lowest coefficient ( $β$ = −3.660). The LASSO regression model for non-depressed diabetes patients had 96.0% accuracy and 89.3% AUC. Amyotrophy had the highest coefficient ( $β$ = 2.712), and age >60 years had the lowest coefficient ( $β$ = −1.415). We found some factors were associated with suicide attempts in depression but were not associated with suicide attempts in non-depression, such as male, Black or African American, Asian, BMI 18.5-24.9 kg/m², T2DM, cerebral infarction, hyperosmolarity, tricyclic Antidepressants, average random glucose 200-300 mg/dl, average cholesterol <200 mg/dl, average cholesterol ≥200 mg/dl, and LDL 130-159 mg/dl (Table 2).

Discussion

To our knowledge, this is the first study that used LASSO regression analysis with the largest dataset to identify associated factors of suicide attempts in people with diabetes. We conducted a literature review to identify candidate risk factors as variables. Then, we performed multivariate and LASSO regression analyses to determine statistically significant risk factors for suicide attempts among people with diabetes using the dataset from the Cerner Real-World Data™. We compared the performances between the multivariate regression model and the LASSO regression model. The accuracy and AUC of those two models were similar. However, the multivariate model had a low precision (1.7%) and F1 score (3.3%). The LASSO model had 87.8% precision related to a low false positive rate and achieved an 87.4% F1 score or weighted average of precision and recall. Overall, we found that the LASSO model performed better than the multivariate regression model, which suggests that LASSO was a better approach for this study. After analyzing data using LASSO regression, we discovered several known and unknown risk factors for suicide attempts in people with diabetes. We also identified risk factors for suicide attempts among each subgroup: males with diabetes, females with diabetes, T1DM, T2DM, people with diabetes and depression, and people with diabetes and non-depression.

Significant findings

We discovered additional risk factors for suicide attempts among people with diabetes that were not considered as known risk factors in the literature, such as being an American Indian or Alaska Native ( $β$ = 0.637), antihistamines ( $β$ = 0.528), atypical agents ( $β$ = 0.704), and benzodiazepines ( $β$ = 0.784). To our knowledge, no study reported statistics of suicidal behaviors and diabetes in American Indians or Alaska Natives people. However, a report by the US Department of Health and Human Services, Office of Minority Health stated that American Indians or Alaska Natives were about three times more likely to be diagnosed with diabetes and were 1.2 times more likely to complete suicide than White.^34,35 In our study, antihistamine drug classes of anti-anxiety medications were associated with suicide attempts in people with diabetes. Drug Information Database reported that hydroxyzine could cause mental depression and suicidal behaviors.³⁶ On the other hand, the study by Gibbons et al. reported that hydroxyzine (OR: 0.53, 95% CI: 0.42-0.65) decreased the risk of suicidal events.³⁷ Further studies on hydroxyzine interactions with depression and suicidal behaviors are needed. Our study found atypical antidepressants increased the risk of suicide attempts among people with diabetes in this study. In general, most antidepressants, including atypical antidepressants, are safe and effectively reduce suicide risk in adults. However, antidepressants can increase the risk of suicidal behaviors in children, adolescents, and young adults, especially at the beginning of treatment.^38,39 Further studies are needed to confirm the relationship between atypical antidepressants and suicide in diabetes and determine drug interactions between atypical antidepressants and diabetes medications.

Amyotrophy and gender

Diabetic amyotrophy is a rare diabetes complication (<1%) characterized by weakness and wasting of the muscles in the legs and hips, as well as pain and numbness in these areas.⁴⁰ Our study found that amyotrophy had a negative association with suicide attempts in males with diabetes. In contrast, it had a positive association with suicide attempts in females with diabetes. To our knowledge, no study has identified the relationship between diabetes amyotrophy and suicide. Erlangsen et al. studied more than 7 million individuals to determine the relationship between neurological disorders and suicide. The study reported that individuals diagnosed with neurological disorders had a 1.8 times higher incidence rate of suicide than those without neurological disorders, and the incidence rate increased to 4.9 times for individuals with amyotrophic lateral sclerosis.⁴¹

MAOI and diabetes

When comparing T1DM and T2DM subgroups, we found that T1DM patients who received MAOI had a negative association with suicide attempts. Emory and Mizrahi studied MAOI in patients with T1DM and depression.⁴² They explained that MAOI in T1DM patients increases dopamine levels, insulin signaling, and catecholamines levels, which leads to reduced HbA1c and blood sugar. MAOI also increases serotonin levels, which consequently leads to improved depression symptoms. This suggests T1DM patients who received MAOI may have decreased risk of suicide attempts.

Age and depression

Depression is known as a major risk factor for suicide.^13-18 Therefore, we decided to compare depressed and non-depressed diabetes patients with suicide attempts. We found that patients with diabetes aged less than 20 had the highest association with suicide attempts for both groups. Gómez-Peralta et al. found that younger patients with diabetes had three to four times increased odds of attempted suicide over the older population (44.75 ± 14.01, p = 0.001).⁹ Robinson et al. also supported that patients with diabetes aged between 15 and 25 had a three-fold increase in attempted suicide compared to non-diabetic people.⁴³ The reasons may be explained by the challenges of managing diabetes at a younger age. In addition, factors such as feeling like a burden to family members or caregivers, the daily requirement for insulin injections, blood glucose monitoring, and dietary restrictions may contribute to suicidal behaviors.^43-45 However, further research is needed to explore the role of age in suicidal behaviors in people with diabetes.

Limitations

Our study has some limitations that should be noted. Firstly, as a cross-sectional study, our findings cannot establish a cause-and-effect relationship between risk factors and suicide attempts among those with diabetes. Secondly, the majority of participants were White and non-Hispanic, which means that the findings may have limited generalizability. Thirdly, electronic health records were used as the data source, which may contain incomplete, inconsistent, or inaccurate information. However, we controlled data quality by removing risk factors with high amounts of missing or outlier values. Finally, the proportion of suicide attempts among diabetes patients in our study was lower than what has been reported in other studies. However, the sample size of our study has sufficient power (>95%) to detect a significant difference between diabetes patients with and without suicide attempts.

Conclusion

To summarize, our study aimed to identify the factors associated with suicide attempts in people with diabetes by analyzing a large dataset using LASSO regression. We uncovered several associated factors, including new findings such as being an American Indian or Alaska Native, antihistamines, atypical agents, and benzodiazepines. Additionally, amyotrophy was associated with suicide attempts in both males and females with diabetes. Furthermore, our study found that T1DM who received MAOI medication had a negative association with suicide attempts. Lastly, patients with diabetes aged less than 20 had the highest association with suicide attempts for both patients with and without depression. Our findings can be utilized by healthcare providers as clinical decision support to identify high-risk groups for suicide attempts among diabetes patients and provide necessary preventive measures during outpatient clinic visits. However, it should be noted that our study was cross-sectional and could not establish causality. Further research in the form of cohort studies is needed to study cause-and-effect relationships between risk factors and suicide attempts in diabetes patients.

Footnotes

Authors’ Contributions

Ploypun Narindrarangkura: Conceptualization, Methodology, Software, Data curation, Writing- Original draft preparation, Visualization, Investigation. Ploypun Narindrarangkura: Supervision. Patricia E. Alafaireet, Uzma Khan, Kim Min Soon: Writing- Reviewing and Editing.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Agency for Healthcare Research and Quality [R21HS028032]; and the National Institute of Diabetes and Digestive and Kidney Disorders [P30DK092950].

ORCID iDs

Ploypun Narindrarangkura

Uzma Khan

References

Ali

Stone

Peters

Davies

Khunti

. The prevalence of co-morbid depression in adults with Type 2 diabetes: A systematic review and meta-analysis. Diabet Med J Br Diabet Assoc. 2006;23:1165-1173.

Anderson

Freedland

Clouse

Lustman

. The prevalence of comorbid depression in adults with diabetes: A meta-analysis. Diabetes Care. 2001;24:1069-1078.

Barker

Ford

Zhang

Strine

Mokdad

. Diabetes and anxiety in US adults: Findings from the 2006 behavioral risk factor surveillance system. Diabet Med J Br Diabet Assoc. 2008;25:878-881.

Ford

Zhao

, et al. Association between diagnosed diabetes and serious psychological distress among U.S. adults: The Behavioral Risk Factor Surveillance System, 2007. Int J Public Health. 2009;54(Suppl 1):43-51.

Roy

Janal

. Suicide attempts and ideation in African-American type 1 diabetic patients. Psychiatr Res. 2010;179:53-56.

Ribeiro

Huang

Fox

Franklin

. Depression and hopelessness as risk factors for suicide ideation, attempts and death: Meta-analysis of longitudinal studies. Br J Psychiatry. 2018;212:279-286.

Hawton

Casañas i Comabella

Haw

Saunders

. Risk factors for suicide in individuals with depression: A systematic review. J Affect Disord. 2013;147:17-28.

Myers

Grannemann

Lingvay

Trivedi

. Brief report: Depression and history of suicide attempts in adults with new-onset Type 2 Diabetes. Psychoneuroendocrinology. 2013;38:2810-2814.

Gómez-Peralta

González-Castro

Fresan

, et al. Risk factors and prevalence of suicide attempt in patients with type 2 diabetes in the Mexican population. Int J Environ Res Public Health. 2018;15:1198.

10.

Radobuljac

Bratina

Battelino

Tomori

. Lifetime prevalence of suicidal and self-injurious behaviors in a representative cohort of Slovenian adolescents with type 1 diabetes. Pediatr Diabetes. 2009;10:424-431.

11.

Chung

Moon

Kim

Min

Kim

Hwang

. Suicidal ideation and suicide attempts among diabetes mellitus: The Korea National Health and Nutrition Examination Survey (KNHANES IV, V) from 2007 to 2012. J Psychosom Res. 2014;77:457-461.

12.

Westling

Ahrén

Träskman-Bendz

Westrin

. High CSF-insulin in violent suicide attempters. Psychiatr Res. 2004;129:249-255.

13.

Bernal

Haro

Bernert

, et al. Risk factors for suicidality in Europe: Results from the ESEMED study. J Affect Disord. 2007;101:27-34.

14.

Sarkar

Balhara

YPS

. Diabetes mellitus and suicide. Indian J Endocrinol Metab. 2014;18:468-474.

15.

Deschênes

Burns

Schmitz

. Comorbid depressive and anxiety symptoms and the risk of type 2 diabetes: Findings from the lifelines cohort study. J Affect Disord. 2018;238:24-31.

16.

Pompili

Forte

Lester

, et al. Suicide risk in type 1 diabetes mellitus: A systematic review. J Psychosom Res. 2014;76:352-360.

17.

Pompili

Lester

Innamorati

, et al. Quality of life and suicide risk in patients with diabetes mellitus. Psychosomatics. 2009;50:16-23.

18.

C-S

Hsu

L-Y

Wang

S-H

. Association of depression and diabetes complications and mortality: A population-based cohort study. Epidemiol Psychiatr Sci. 2020;29:e96.

19.

Sperandei

. Understanding logistic regression analysis. Biochem Med. 2014;24:12-18.

20.

Friedman

Hastie

Tibshirani

. Regularization paths for generalized linear models via coordinate descent. J Stat Softw. 2010;33:1-22.

21.

Tibshirani

. Regression Shrinkage and selection via the Lasso. J R Stat Soc Ser B Methodol. 1996;58:267-288.

22.

Muthukrishnan

Rohini

. LASSO: A feature selection technique in predictive modeling for machine learning. 2016 IEEE International Conference on Advances in Computer Applications (ICACA), Coimbatore, India, 24-24 October; 2016:18-20.

23.

Plows

Stanley

Baker

Reynolds

Vickers

. The pathophysiology of gestational diabetes mellitus. Int J Mol Sci. 2018;19:3342.

24.

Blehar

Spong

Grady

Goldkind

Sahin

Clayton

. Enrolling pregnant women: Issues in clinical research. Womens Health Issues Off Publ Jacobs Inst Womens Health. 2013;23:e39-e45.

25.

Ehwerhemuepha

Carlson

Moog

, et al. Cerner real-world data (CRWD) - A de-identified multicenter electronic health records database. Data Brief. 2022;42:108120.

26.

Ehwerhemuepha

Gasperino

Bischoff

Taraman

Chang

Feaster

. HealtheDataLab - a cloud computing solution for data science and advanced analytics in healthcare with application to predicting multi-center pediatric readmissions. BMC Med Inform Decis Mak. 2020;20:115.

27.

Project Jupyter , https://www.jupyter.org (Accessed 11 December, 2021).

28.

Apache Spark . Apache Spark^TM - Unified Engine for large-scale data analytics; 2021. https://spark.apache.org/. Accessed December 11, 2021.

29.

Rengaraju

. Handling missing values. WiDS Mysore; 2020. https://medium.com/wids-mysore/handling-missing-values-82ce096c0cef. Accessed February 14, 2022.

30.

van Buuren

Groothuis-Oudshoornmice

Multivariate Imputation by Chained Equations in R. J Stat Soft 2011;45(3):1-67. https://doi.org/10.18637/jss.v045.i03

31.

Lunardon

Menardi

Torelli

. ROSE: Random Over-Sampling Examples; 2021. https://CRAN.R-project.org/package=ROSE. Accessed October 21, 2022.

32.

Kautzky-Willer

Harreiter

Pacini

. Sex and gender differences in risk, pathophysiology and complications of type 2 diabetes mellitus. Endocr Rev. 2016;37:278-316.

33.

Wang

Shi

Zhang

. Suicide risk in patients with diabetes: A systematic review and meta-analysis. Eur J Endocrinol 2017;177(4):R169-R181. doi:10.1530/EJE-16-0952

34.

Office of Minority Health Resource Center . Diabetes and American Indians/Alaska Natives; 2021. https://minorityhealth.hhs.gov/omh/browse.aspx?lvl=4&lvlid=33. Accessed April 7, 2022.

35.

Office of Minority Health Resource Center . Mental and Behavioral Health - American Indians/Alaska Natives; 2020. https://minorityhealth.hhs.gov/omh/browse.aspx?lvl=4&lvlid=39. Accessed April 7, 2022.

36.

Hydroxyzine disease interactions . Drugs.com. https://www.drugs.com/disease-interactions/hydroxyzine.html (2022, Accessed 28 December 2022).

37.

Gibbons

Hur

Lavigne

Wang

Mann

. Medications and suicide: High Dimensional Empirical Bayes Screening (iDEAS). Harvard Data Science Review. 2019. Epub ahead of print 1 November 2019. doi:10.1162/99608f92.6fdaa9de

38.

Pompili

Serafini

Innamorati

, et al. Antidepressants and suicide risk: A comprehensive overview. Pharmaceuticals. 2010;3:2861-2883.

39.

Spielmans

Spence-Sing

Parry

. Duty to Warn: antidepressant black box suicidality warning is empirically justified. Front Psychiatry. 2020;11:18.

40.

Garland

. Diabetic amyotrophy. Br Med J. 1955;2:1287-1290.

41.

Erlangsen

Stenager

Conwell

, et al. Association between neurological disorders and death by suicide in denmark. JAMA. 2020;323:444-454.

42.

Emory

Mizrahi

. Monoamine oxidase inhibition in a patient with Type 1 diabetes and depression. J Diabetes Sci Technol. 2016;10:1203-1204.

43.

Robinson

M-E

Simard

Larocque

Shah

Nakhla

Rahme

. Risk of psychiatric disorders and suicide attempts in emerging adults with diabetes. Diabetes Care. 2020;43:484-486.

44.

Majidi

O’Donnell

Stanek

Youngkin

Gomer

Driscoll

. Suicide risk assessment in youth and young adults with Type 1 diabetes. Diabetes Care. 2020;43:343-348.

45.

Markowitz

Volkening

Butler

Laffel

LMB

. Youth-perceived burden of Type 1 diabetes: Problem Areas in Diabetes Survey–Pediatric Version (PAID-Peds). J Diabetes Sci Technol. 2015;9:1080-1085.