Predictive models and determinants of mortality among T2DM patients in a tertiary hospital in Ghana, how do Machine Learning Techniques perform?

doi:10.21203/rs.3.rs-4359019/v1

Background: The increasing prevalence of type 2 diabetes mellitus (T2DM) in lower and middle – income countries calls for preventive public health interventions. Studies from Africa including those from Ghana, consistently reveal high T2DM-related mortality rates. While previous research in the Ho municipality has primarily examined risk factors, comorbidity, and quality of life of T2DM patients, this study specifically investigated mortality predictors among these patients.

Method: The study was retrospective involving medical records of T2DM patients. Data extracted were analyzed using Stata version 16.0 and Python 3.6.1 programming language. Both descriptive and inferential statistics were done to describe and build predictive models respectively. The performance of machine learning (ML) techniques such as support vector machine (SVM), decision tree, k nearest neighbor (kNN) and logistic regression were evaluated using the best-fitting predictive model of T2DM mortality.

Results: Out of the 328 participants, 183(55.79%) were females. An 11.28% mortality was recorded. A 100% mortality was recorded among the T2DM patients with sepsis (p-value = 0.012). T2DM patients were 3.83 times as likely to die [AOR = 3.83; 95% CI: (1.53-9.61)] if they had nephropathy compared to T2DM patients without nephropathy (p-value = 0.004). The full model which included sociodemographic characteristics, family history, lifestyle variables and complications of T2DM had the best prediction of T2DM mortality outcome (ROC = 72.97%). The accuracy for (test and train datasets) were as follows: (90% and 90%), (100% and 100%), (90% and 90%) and (90% and 88%) respectively for the various classification techniques: logistic regression, Decision tree classifier, kNN classifier and SVM.

Conclusion: This study found that all patients with sepsis died. Nephropathy was the identified significant predictor of T2DM mortality. Decision tree classifier provided the best classifying potential.

Type 2 Diabetes Mellitus

Predictors

Mortality

Machine Learning Techniques

Model

The rise in Non-Communicable Diseases (NCDs), notably Type 2 Diabetes Mellitus (T2DM), in developing nations coupled with the challenges faced by healthcare systems in managing this growing burden (1), highlights the urgent need for preventive public health measures (2). This concerning trend of rising T2DM cases is especially evident among two specific demographic groups: older adults and obese young individuals (3–5). The reasons behind this trend could be biological or closely linked to shifts in lifestyle and economic improvements (6, 7). For instance, in Ghana, the reported prevalence of T2DM stands at 3.95% among individuals aged 50 years or older (6).

It is crucial to emphasize that T2DM, along with its associated complications such as cardiovascular diseases (CVDs) and chronic kidney disease (CKD), significantly contribute to mortality rates (8–10). In fact, T2DM alone ranks ninth among the leading causes of death worldwide, resulting in over 1 million deaths annually (2). This burden is mirrored in Africa, as exemplified by Nigeria, where mortality and case fatality rates of T2DM are reported at 30.2 per 100,000 population and 22.0%, respectively (11). A study conducted in Ghana further underscores this trend, revealing that over a 31-year period (1983–2012), hospitalized case fatality rates due to diabetic conditions surged from 7.6 per 1000 deaths to 30 per 1000 deaths (12). The authors of the study also noted an average of 18.5% of deaths occurring approximately every 28 days (12).

In the Ho Municipality, a significant number of annual new diabetes cases, totaling 511, were reported in 2021 (13). While efforts are made to improve the well-being of individuals living with T2DM in the Municipality, the research focus has primarily centered on addressing various associated risk factors, comorbidity, and enhancing their quality of life (14–16). However, a noticeable gap in the existing body of knowledge pertains to the exploration of predictors of mortality among T2DM patients. To bridge this knowledge gap, the current study aimed to identify the predictors of mortality among T2DM patients receiving care at the Ho Teaching Hospital in Ghana. By pinpointing these factors, it becomes possible to target risk reduction strategies, ultimately leading to a reduction in mortality rates among T2DM patients. This aligns with the second objective of the National Policy for the Prevention and Control of Chronic Non-Communicable Diseases (NCDs) in Ghana, which aims to minimize exposure to risk factors contributing to NCDs, including T2DM (17). Importantly, it's worth noting that significant advancements in preventive health have been achieved through the application of ML techniques (18). Consequently, this study also incorporates an evaluation of the predictive potential of selected ML techniques, aiming to harness these innovative methods for enhancing T2DM care and management.

Study Design

The study retrospectively analyzed the medical records of Type 2 Diabetes Mellitus (T2DM) patients treated at the Ho Teaching Hospital (HTH) between January 2017 and November 2022.

Study area

The study was carried out at HTH. The hospital facility is the main referral center in the Volta Region. HTH is the fifth public Teaching Hospital in Ghana and serves the needs of the region and beyond. It has seven directorates (Medical Affairs, Administration & Support Services, Nursing Administration, Human Resources, Research, Innovation, Planning, Monitoring and Evaluation, Finance and Pharmacy) (19). The Hospital has over 300-bed capacity to cater for the health needs of patients (19). Among the clinical department in the facility include, Internal Medicine, Surgical, Obstetrics & Gyaenacology, Child Health and Public Health (19). The diabetic services for pateints include consulting with the dietician, and visiting the general clinic. Specific laboratory investigations such as FBG, BMI, lipid profile, urine glucose, kidney function test, and liver function test are carried out in addition to checking for compliance with medication, monitoring co-morbidity and complication.

Study population

The study population comprised all the accessible medical records of T2DM patients aged 18 years and older who received healthcare services at HTH between January 2017 and November 2022.

Inclusion and Exclusion Criteria

Inclusion criteria

The electronic and manual medical records of in-patients aged 18 years and above who had complete sociodemographic characteristics data as well as lifestyle variables, complications of diabetes and mortality outcome within the stipulated period for the study (January 2017 to November 2022) were included into the study.

Exclusion criteria

Patients with Type 1 diabetes, T2DM out – patients as well as the T2DM patients whose data on the lifestyle variables, complications of diabetes and mortality outcome could not be found in the medical records (electronic and manual) were excluded.

Sample Size

The sample size for the study was determined by using the Cochran formula (Cochran, 1977) for cross-sectional studies: $n=\frac{{P\left(1-P\right) x (Z\alpha /2)}^{2}}{{e}^{2}}$,

Where:

n is the estimated sample size,

Z_α/2 is the reliability coefficient (1.96 at the 95% confidence level),

p is the national mortality = 3.39% and

e is margin of error allowable for this study (5%).

By substituting the figures into the formula,

$$n=\frac{{0.0339*\left(1-0.0339\right) x \left(1.96\right)}^{2}}{{0.05}^{2}}=50.33$$

A complete enumeration of the study population was employed where all medical records (electronic and manual) of T2DM in – patients, aged 18 years and above who accessed health care at the HTH from January 2017 to November 2022 was done resulting in a total of 328 samples (241 electronic and 87 manual records) used for the study.

Study variables

The dependent variable of the study was mortality outcome among T2DM. The independent variables included sociodemographic characteristics, family history of disease(s) lifestyle variables and complications of diabetes.

Data retrieval and management

Data was retrieved using a Microsoft (MS) Excels version 2016. Data were extracted from the electronic and manual patient folders. A data extraction sheet was used to capture data on sociodemographic characteristics (age, sex, marital status, family history, educational level, occupation and location), lifestyle variables (smoking and alcohol intake), and haemodynamic variables (SBP and DBP). The resulting data collated was coded and cleaned in MS Excel and password protected. Mortality outcome was coded 1 and 0 (1 = Dead and 0 = Alive).

Data Analysis

Data extracted were entered into MS Excel version 2016 and analyzed using Stata version 16.0 and Python 3.6.1 programming language. Quantitative variables were presented as Mean ± SD for those that were parametric and median (Interquartile Range) for those that were non-parametric. Frequencies and percentages were used to summarize categorical variables. To understand the associations of independent variables with the outcome, Chi-square test was done for categorical variables while independent t – test (for parametric variables). Univariable logistic regression was also used to obtain the crude strength of association between the mortality and the independent variables. Multiple logistic regression was done to obtain the adjusted odds ratio. A p-value of 0.05 was considered statistically significant. The predictive models were evaluated using the area under the ROC curve. Multicollinearity was tested using generalized variance inflation factor for logistic regression model. The best-performing regression model was evaluated using scikit learn ML module in Python. The dataset was divided into test (70%) and train (30%). Four classifiers, logistic regression, decision tree, k nearest neighbour (kNN) and Support Vector Machine (SVM) were used as learners for the classification.

Ethical Consideration

This study's protocol obtained ethical approval from the Research and Ethics Committee of the Ho Teaching Hospital, identified by Protocol ID No: HTH-REC (20) FC_2022. Additionally, permission was granted by the facility's records department prior to commencing data collection.

The study recorded a mortality of 11.28% among patients diagnosed with Type 2 Diabetes at the Ho Teaching Hospital from 2017 to 2022 (Fig. 1).

Table 1 shows the sociodemographic characteristics of T2DM patients seeking health care at the Ho Teaching Hospital. The study observed a female preponderance of 183(55.79%) and the average age of participants was 58.61 ± 14.64 years. More than half were married 221(67.38%), located in urban settlements 187(57.01%) and formally employed 174(53.05%). One hundred and thirty-two of the T2DM patients representing 40.24% (95% CI: 35.05% − 45.67%) had primary education while the minority did not have any form of formal education 45(13.72%). The patients also had family history of diabetes 54(16.48%), cardiovascular diseases 67(20.43%), asthma 5(1.52%) and both cardiovascular diseases and diabetes 34(10.37%) as well as cardiovascular diseases, diabetes and asthma 1(0.35%). The lifestyle characteristics of participants also showed that 65(19.82%) and 20(6.1%) were current alcoholic and smoker respectively.

Table 1

Sociodemographic characteristics and disease family history of T2DM patients seeking health care at the Ho Teaching Hospitals
Variable	n(%)
*Age(years)	58.61 ± 14.64
Sex
Male	145(44.21)
Female	183(55.79)
Marital Status
Single	73(22.26)
Married	221(67.38)
Widowed	28(8.54)
Divorced	6(1.83)
Location
Urban	187(57.01)
Rural	141(42.99)
Occupation
Unemployed	49(14.94)
Informal	174(53.05)
Formal	56(17.07)
Retired	49(14.94)
Educational Level
None	45(13.72)
Primary	132(40.24)
Secondary	55(16.77)
Tertiary	96(29.27)
Family history
Diabetes	54(16.48)
Cardiovascular diseases	67(20.43)
Asthma	5(1.52)
Cardiovascular diseases & Diabetes	34(10.37)
Cardiovascular diseases, Diabetes & Asthma	1(0.30)
Lifestyle History
Drink Alcohol	65(19.82)
Smoke	20(6.10)

Data are presented as frequencies and percentages. * is presented as mean ± Standard Deviation

Table 2 shows the chi-square test of association and binary logistic regression of sociodemographic factors and family history of disease with mortality outcome. None of the sociodemographic factors and family history of disease were significantly associated with mortality.

Table 2

Association of sociodemographic factors and family history of disease with mortality among T2DM patients seeking health care at the Ho Teaching Hospital
Variables	Mortality Outcome		Chi-square p-value	cOR(95%CI)	p-value
Variables	Alive	Dead	Chi-square p-value	cOR(95%CI)	p-value
Age*	58.44 ± 14.62	59.95 ± 14.97	0.557	1.00(0.98–1.03)	0.555
Sex
Male	124(85.52)	21(14.48)	0.103	Ref
Female	167(91.26)	16(8.74)		0.54(0.28–1.13)	0.106
Marital Status
Single	62(84.93)	11(15.07)	0.294	Ref
Married	200(90.50)	21(9.50)		0.59(0.27–1.30)	0.189
Widowed	23(82.14)	5(17.86)		1.23(0.38–3.91)	0.731
Divorced	6(100.00)	0(0.00)		-
Location
Urban	164(88.17)	22(11.83)	0.587	Ref
Rural	127(90.07)	14(9.93)		0.78(0.39–1.59)	0.502
Occupation
Unemployed	43(87.76)	6(12.24)	0.894	Ref
Informal	154(88.51)	20(11.49)		0.93(0.35–2.46)	0.885
Formal	49(87.50)	7(12.50)		1.02(0.32–3.28)	0.968
Retired	45(91.84)	4(8.16)		0.64(0.17–2.41)	0.507
Educational Level
None	38(84.44)	7(15.56)	0.163	Ref
Primary	120(90.91)	12(9.09)		0.54(0.20–1.48)	0.232
Secondary	52(94.55)	3(5.45)		0.31(0.08–1.29)	0.108
Tertiary	81(84.38)	15(15.63)		1.01(0.38–2.67)	0.992
Family history^	Total = 291	Total = 37
Diabetes	48(16.49)	6(16.22)	0.966	0.98(0.39–2.48)	0.966
Cardiovascular diseases (CVD)	60(20.62)	7(18.92)	0.809	0.90(0.38–2.14)	0.809
Asthma	4(1.37)	1(2.70)	0.452	2.00(0.22–18.32)	0.542
CVD & Diabetes	29(9.97)	5(13.51)	0.505	1.41(0.51–3.91)	0.507
CVD, Diabetes & Asthma	0(0.00)	1(2.70)	0.113	-
Lifestyle History^	Total = 291	Total = 37
Current Alcoholic	58(19.93)	7(18.92)	0.884	0.94(0.39–2.24)	0.884
Current Smoker	16(5.50)	4(10.81)	0.203	2.08(0.66–6.60)	0.212

^ Proportions are calculated column-wise. * presented as mean ± SD. P-value significant at < 0.05

The chi-square test of association and univariate binary logistic regression of T2DM complications and comorbidities with mortality outcome showed that among the T2DM patients’ complications and comorbidity, nephropathy, as a complication and sepsis were significantly associated with diabetic mortality. Higher proportions of those who died presented with nephropathy 25% compared to those without nephropathy 9.15% (p-value = 0.002). The crude odds ratio shows that those having nephropathy had a threefold increased odds of mortality compared to those without [cOR = 3.31, 95%CI (1.50–7.31); p-value = 0.003]. The only two patients who presented with sepsis in the study died (p-value = 0.012) [Table 3].

Table 3

Association of diabetic complications and comorbidities with mortality among study participants
Variables	Mortality Outcome		Chi-square p-value	cOR(95%CI)	p-value
Variables	Alive	Dead	Chi-square p-value	cOR(95%CI)	p-value
Cardiovascular disorders
Absent	227(87.98)	31(12.02)	0.419	Ref
Present	64(91.43)	6(8.57)		0.68(0.27–1.72)	0.421
Nephropathy
Absent	258(90.85)	26(9.15)	0.002	Ref
Present	33(75.00)	11(25.00)		3.31(1.50–7.31)	0.003
Neuropathy
Absent	290(88.96)	36(11.04)	0.085	Ref
Present	1(50.00)	1(50.00)		8.06(0.49-131.59)	0.143
Foot Ulcer
Absent	279(88.57)	36(11.43)	1.000	Ref
Present	12(92.31)	1(7.69)		0.65(0.08–5.11)	0.679
Diabetic Ketoacidosis
Absent	240(87.91)	33(12.09)	0.303	Ref
Present	51(92.73)	4(7.27)		0.57(0.19–1.68)	0.309
Hyperosmolarity coma
Absent	284(88.75)	36(11.25)	1.000	Ref
Present	7(87.50)	1(12.50)		1.13(0.13–9.42)	0.912
Hyperosmolarity without coma
Absent	282(88.96)	35(11.04)	0.358	Ref
Present	9(81.82)	2(18.18)		1.79(0.37–8.62)	0.468
Other conditions
Skin Infection
Absent	286(88.54)	37(11.46)	1.000
Present	5(100.00)	0(0.00)		-
Sickle Cell
Absent	291(88.99)	36(11.01)	0.113
Present	0(0.00)	1(100.00)		-
Sepsis
Absent	291(89.26)	35(10.74)	0.012	-
Present	0(0.00)	2(100.00)
Urinary Tract Infection
Absent	282(88.40)	37(11.60)	0.605
Present	9(100.00)	0(0.00)		-
Pneumonia
Absent	286(88.82)	36(11.18)	0.515	Ref
Present	5(83.33)	1(16.67)		1.59(0.18–13.98)	0.676
P-value < 0.05 is considered significant

After adjusting for sociodemographic and all other factors, the only statistically significant factor contributing to mortality among the T2DM patients seeking health care at the Ho Teaching Hospital was nephropathy. T2DM patients with nephropathy had 3.83-fold odds of death [95% CI: (1.53–9.61)] compared to T2DM patients without nephropathy. This was statistically significant at a p-value of 0.004 (Table 4).

Table 4

Predictive models for predicting mortality among T2DM patients seeking health care at the Ho Teaching Hospital
Variables	Model 1		Model 2		Model 3
Variables	aOR(95%CI)	p-value	aOR(95%CI)	p-value	aOR(95%CI)	p-value
Family history
Diabetes	1.17(0.37–3.73)	0.779	1.23(0.38–3.99)	0.728	1.02(0.29–3.56)	0.977
Cardiovascular diseases	0.84(0.28–2.55)	0.761	0.83(0.27–2.51)	0.741	0.90(0.26–3.05)	0.909
Asthma	2.13(0.17–26.84)	0.559	2.38(0.20–28.60)	0.493	3.13(0.22–44.83)	0.436
Lifestyle History
Current Alcoholic			0.60(0.18–1.99)	0.407	0.74(0.21–2.62)	0.612
Current Smoker			2.67(0.58–12.25)	0.207	1.87(0.37–9.50)	0.448
Complications
Cardiovascular diseases					0.79(0.27–2.28)	0.665
Nephropathy					3.83(1.53–9.61)	0.004
Neuropathy					8.51(0.39-187.28)	0.175
Foot Ulcer					0.51(0.05–4.75)	0.553
Diabetic Ketoacidosis					0.64(0.19–2.17)	0.472
Hyperosmolarity coma					1.38(0.14–14.11)	0.784
Hyperosmolarity without coma					1.64(0.26–10.25)	0.600
Pneumonia					1.41(0.11–17.38)	0.788

aOR = Adjusted Odds ratio controlling for sociodemographic variables. P-value significant at < 0.05

As seen in Fig. 2, the area under the ROC curve was highest for Model 3 (ROC = 72.97%) among the three models (Model 1 = 67.03% and Model 2 = 67.85%), making it the preferred model and indicating a good predictive ability of the fitted model to predict mortality among T2DM patients.

Examination of the GVIF values for the preferred model (model 3) showed that all the GVIF values were far less than the cut-off value of 10 and the mean GVIF was 1.23 which is less than 6, indicating no multicollinearity in the model (Table 5).

Table 5

Test of multicollinearity using generalised variance inflation factor for logistic regression model (model 3)
Variables	GVIF	Degree of freedom (DF)	GVIF^(1/2*DF)
Age	1.16		0.862069
Sex	1.27	1	1.13
Marital Status	1.17	1	1.08
Location	1.11	1	1.05
Occupation	1.75	3	2.32
Educational Level	1.64	3	2.1
Family history
Diabetes	1.42	1	1.19
Cardiovascular diseases	1.44	1	1.2
Asthma	1.1	1	1.05
Lifestyle History
Drink Alcohol	1.41	1	1.19
Smoke	1.36	1	1.17
Complications
Cardiovascular diseases	1.09	1	1.04
Nephropathy	1.12	1	1.06
Neuropathy	1.05	1	1.02
Foot Ulcer	1.04	1	1.02
Diabetic Ketoacidosis	1.11	1	1.05
Hyperosmolarity coma	1.05	1	1.02
Hyperosmolarity without coma	1.04	1	1.02
Pneumonia	1.08	1	1.04

The performance of four machine learning technique: logistic regression, decision tree, kNN, and SVM were evaluated using the best – performing predictive model. The accuracy results for (test and train datasets) were as follows: (90% and 90%), (100% and 100%), (90% and 90%) and (90% and 88%) respectively for the various classification techniques: logistic regression, Decision tree classifier, kNN classifier and SVM. Additionally, the precision, recall and F1 score for decision tree were almost 1. Thus, making the decision tree the best classifier of the four (Table 6).

Table 6

Performance of Decision Tree, kNN, logistic regression, and SVM in predicting mortality outcome among T2DM patients
Classifier		Train	Test
Decision Tree
Accuracy		1.00	1.00
Precision	Alive	1.00	1.00
Precision	Dead	1.00	0.94
Recall	Alive	1.00	0.99
Recall	Dead	1.00	1.00
F1 score	Alive	1.00	1.00
F1 score	Dead	1.00	0.97
kNN
Accuracy		0.90	0.90
Precision	Alive	0.99	1.00
Precision	Dead	0.20	0.06
Recall	Alive	0.90	0.90
Recall	Dead	0.80	1.00
F1 score	Alive	0.94	0.95
F1 score	Dead	0.32	0.11
Logistic Regression
Accuracy		0.90	0.90
Precision	Alive	1.00	1.00
Precision	Dead	0.15	0.00
Recall	Alive	0.89	0.90
Recall	Dead	1.00	0.00
F1 score	Alive	0.94	0.95
F1 score	Dead	0.26	0.00
SVM
Accuracy		0.88	0.90
Precision	Alive	1.00	1.00
Precision	Dead	0.00	0.00
Recall	Alive	0.88	0.90
Recall	Dead	0.00	0.00
F1 score	Alive	0.94	0.95
F1 score	Dead	0.00	0.00
kNN: K nearest Neighbor, SVM: Support Vector Machine

An overall T2DM mortality of 37(11.28%) was observed in this study. A study in Denmark by Reinhard, Lajer (20) reported a higher rate of T2DM (68%) while a lower rate of 1.93% was reported by Mulnier, Seaman (21) in the UK. This variation in the study outcomes may be due to differences in geographical location, study design or sample size. While this study was a cross-sectional study using secondary data with a sample size of 328; the studies by Reinhard, Lajer (20) and Mulnier, Seaman (21) were cohort studies and employed sample sizes of 283 and 44,230 T2DM records respectively. It's essential to recognize that cross-sectional retrospective studies face limitations in establishing causality, primarily because they cannot accurately determine the temporal sequence of events. Consequently, care should be taken in generalizing this findings to different time periods, particularly if there have been shifts in exposures, outcomes, or other pertinent factors over time. Additionally, although our study focused on T2DM cause-specific mortality, a national study in Ghana found a slightly higher mortality rate for chronic NCDs to be 20 deaths per 100 admissions in all health institutions making it the leading cause of death (22). Emphasizing preventive care is essential in reducing mortality among T2DM patients. This includes screening for and early management of complications such as diabetic retinopathy, nephropathy, neuropathy, cardiovascular disease, and foot ulcers. By doing so, it becomes possible to reduce T2DM-specific mortality rates within the study jurisdiction.

Consistent with accumulated evidence supporting diabetic complication and comorbidity with mortality (23–25), this study finds nephropathy and sepsis to be significantly associated with T2DM mortality. A 100 percent mortality outcome was observed with T2DM patients who presented with sepsis during the period of the study. A retrospective cohort study by Hsieh, Hu (26) also highlighted the influences of sepsis on mortality among T2DM patients. A plausible explanation for this observation in our study could be that individuals with diabetes face an increased risk of succumbing to infectious diseases (27). This heightened susceptibility is linked to weakened immunity resulting from prolonged poor glycemic control (28). Consequently, when sepsis occurs in T2DM patients, it can lead to a fatal outcome.

Another key finding predicting mortality in this study is nephropathy. Nephropathy was reported among 44(13.41%) of the T2DM patients. Twenty five percent of the T2DM patients with nephropathy died as compared to the 9.15% who were without nephropathy. After adjusting for all factors that could potentially confound this association, T2DM patients with nephropathy had 3.83-fold odds of death [95% CI: (1.53–9.61)] compared to T2DM patients without nephropathy. This finding sits well with existing literature. An instance is the study by González-Pérez, Saez (25) who reported that every year, one out of every 20 T2DM patients with Diabetic Kidney Disease (DKD) died. Afkarian, Sachs (29) also found an absolute mortality risk difference with the reference group of 23.4% after adjusting for demographics among diabetics with kidney disease. Considering the adverse mortality outcome associated with nephropathy, a possible waiver or subsidized cost for carrying out kidney function test among T2DM patients should be implemented. Also, annual screening for kidney disease is recommended for all adults with T2DM, starting at the time of diagnosis; however, once kidney disease is detected, the frequency of monitoring may increase, with assessments occurring every 3 to 6 months or as determined by the healthcare provider. This monitory should be done by qualified nephrologist and endocrinologist.

Recognizing the variability in mortality risk among T2DM patients, future research should focus on personalized medicine approaches that tailor treatment strategies to individual patient characteristics, including age, sex, race/ethnicity, duration of diabetes, presence of complications, and socioeconomic status. Also, T2DM patients with nephropathy and sepsis should receive education and support to empower them to actively participate in their care and manage their condition effectively. Such patients should receive aggressive management of these complications including but not limited to optimizing glycemic control, managing blood pressure and lipid levels, addressing underlying renal dysfunction, and implementing infection prevention measures.

Review of existing literature identified sociodemographic characteristics, family history, lifestyle variables and complications of T2DM as independent predictors of mortality among T2DM patients. Three models were evaluated using first the sociodemographic and family history variables, model 2 used sociodemographic, family history and lifestyle variables while the last model used all variables. The findings of this study demonstrate the area under the ROC curve was highest for Model 3 (ROC = 72.97%) among the three models (Model 1 = 67.03% and Model 2 = 67.85%), making it the preferred model and indicating a good predictive ability of the fitted model to predict mortality among T2DM patients. Thus implies that having a holistic medical history of T2DM patients rather than having a single biased view is akin to making better predictions on their health outcome, especially mortality outcome. This finding resonates with the report of Lee, Zhou (30) where a multiparametric model that consisted of different variables of T2DM patients predicted all-cause mortality more accurately.

The current finding is novel in the current study jurisdiction, being the first attempt to investigate the predictive ability of model 3 using ML technique in the study jurisdiction. Consistent with the study by Lee, Zhou (30), predictive model built using ML technique had higher predictive accuracy than the traditional logistic regression model. Of the four learners: logistic regression, Decision tree classifier, kNN classifier and SVM used to train the features of model 3 in this study, decision tree showed the best classifying potential. In a different study to compare the predictive accuracy of SVM, kNN and decision tree algorithms, SVM was found to be the best (31). The dichotomy from the current study finding could be due to the outcomes being predicted in the studies. While this study predicted T2DM mortality outcome, the Wiyono, Wibowo (31) study predicted the performance of students. Nonetheless, studies such as Yue, Xin (32) and Georga, Protopappas (33) used SVM in predicting T2DM. From the current findings, large amount of data is required to better train this model using the decision tree algorithm for mortality prediction among T2DM in the study jurisdiction and the African population at large. Medics and researchers could use this predictive model in the long term to improve the overall mortality outcome of T2DM patients based on their sociodemographic characteristics, family history, lifestyle variables and complications to subsequently initiate early efficient treatment options to avert mortality or reduce its risk.

While this study provides valuable insights, it is essential to acknowledge its limitations. Due to the retrospective design of the study, certain variables, including dietary habits and exercise patterns, which could have yielded valuable lifestyle insights, were not available. Furthermore, it would have been helpful to include data from patients about their current medication as well as other comorbidities. Additionally, the current study was based on a single-site analysis; thus, findings may not be applicable to the whole country.

This study found nephropathy as the significant predictor of T2DM mortality. Also, all patients who were having sepsis during the period under study died. For better prediction of mortality outcome, a holistic assessment of sociodemographic characteristics, family history, lifestyle variables and complications of T2DM is required. Decision tree classifier provided the best classifying potential. Medics and researchers could use this predictive model in the long term to improve the overall mortality outcome of T2DM patients.

ADA

American Diabetes Association

AI

Artificial intelligence

CDC

Center for Disease Control and prevention

CDK

Chronic Kidney Disease

CVD

Cardiovascular Disease

DR

Diabetic Retinopathy

DT

Decision Tree

IDF

International Diabetes Federation

QPSO

Quantum Particle Swarm Optimization

RF

Random Forest

T2DM

Type 2 Diabetes Mellitus

WHO

World Health Organization

SVM

Support Vector Machine.

Ethics approval and consent to participate

This study's protocol obtained ethical approval from the Research and Ethics Committee of the Ho Teaching Hospital, identified by Protocol ID No: HTH-REC (20) FC_2022. Additionally, permission was granted by the facility's records department prior to commencing data collection.

Consent for publication

Not Applicable.

Availability of data and materials

The data used to support the findings of this study are available upon request to the corresponding author.

Competing interests

The authors declare that they do not have any conflicts of interest.

Funding

No funding was received for this study.

Authors' contributions

GEK was supervised by SADO as he conceptualized, developed, gathered, and analyzed the data. GEK wrote the first draft of the manuscript with assistance from SLY. High level review of the manuscript was provided by SLY and SADO. The final paper was approved by all authors.

Acknowledgement

The authors wish to acknowledge the staff of Ho Teaching Hospital, especially Mr. Clement Dason and Mr. Benjamin Amedume for their efforts and assistance during data retrieval.

Authors' information

Department of Medical Laboratory Sciences, School of Allied Health Sciences, University of Health and Allied Sciences, Ho, Ghana

Godsway Edem Kpene

Department of Medical Laboratory Sciences, School of Allied Health Sciences, University of Health and Allied Sciences, Ho, Ghana

Sylvester Yao Lokpo

Department of Public Health Studies. Elon University. Elon, NC. USA

Sandra A. Darfour-Oduro

Kabir A, Karim MN, Islam RM, Romero L, Billah B. Health system readiness for non-communicable diseases at the primary care level: a systematic review. BMJ Open. 2022;12(2):e060387.
Khan MAB, Hashim MJ, King JK, Govender RD, Mustafa H, Al Kaabi J. Epidemiology of Type 2 Diabetes - Global Burden of Disease and Forecasted Trends. J Epidemiol Glob Health. 2020;10(1):107-11.
Ayah R, Joshi MD, Wanjiru R, Njau EK, Otieno CF, Njeru EK, et al. A population-based survey of prevalence of diabetes and correlates in an urban slum community in Nairobi, Kenya. BMC public health. 2013;13(1):1-11.
Samwel Maina G, Benson Williesham M, Miguel San S. Prevalence and determinants of diabetes among older adults in Ghana. BMC public health. 2016;16(1):1-12.
Al-Amiri E, Abdullatif M, Abdulle A, Al-Bitar N, Afandi EZ, Parish M, et al. The prevalence, risk factors, and screening measure for prediabetes and diabetes among Emirati overweight/obese children and adolescents. BMC Public Health. 2015;15(1):1-9.
Gatimu SM, Milimo BW, Sebastian MS. Prevalence and determinants of diabetes among older adults in Ghana. BMC Public Health. 2016;16(1):1174.
de-Graft Aikins A, Addo J, Ofei F, Bosu W, Agyemang C. Ghana's burden of chronic non-communicable diseases: future directions in research, practice and policy. Ghana Med J. 2012;46(2 Suppl):1-3.
McEwen LN, Karter AJ, Waitzfelder BE, Crosson JC, Marrero DG, Mangione CM, et al. Predictors of mortality over 8 years in type 2 diabetic patients: Translating Research Into Action for Diabetes (TRIAD). Diabetes care. 2012;35(6):1301-9.
Mayyas FA, Ibrahim KS. Predictors of mortality among patients with type 2 diabetes in Jordan. BMC Endocrine Disorders. 2021;21(1):200.
Zhou JJ, Koska J, Bahn G, Reaven P. Glycaemic variation is a predictor of all-cause mortality in the Veteran Affairs Diabetes Trial. Diabetes Vascular Disease Research. 2019;16(2):178-85.
Adeloye D, Ige JO, Aderemi AV, Adeleye N, Amoo EO, Auta A, et al. Estimating the prevalence, hospitalisation and mortality from type 2 diabetes mellitus in Nigeria: a systematic review and meta-analysis. BMJ Open. 2017;7(5):e015424.
Sarfo-Kantanka O, Sarfo FS, Oparebea Ansah E, Eghan B, Ayisi-Boateng NK, Acheamfour-Akowuah E. Secular Trends in Admissions and Mortality Rates from Diabetes Mellitus in the Central Belt of Ghana: A 31-Year Review. PLoS One. 2016;11(11):e0165905.
HMHD. Report on T2DM cases 2021.
Osei-Yeboah J, Owiredu W, Norgbe G, Obirikorang C, Lokpo S, Ashigbi E, et al. Physical activity pattern and its association with glycaemic and blood pressure control among people living with diabetes (PLWD) in the Ho municipality, Ghana. Ethiopian journal of health sciences. 2019;29(1).
Osei-Yeboah J, Lokpo SY, Owiredu WK, Johnson BB, Orish VN, Botchway F, et al. Medication Adherence and its Association with Glycaemic Control, Blood Pressure Control, Glycosuria and Proteinuria Among People Living With Diabetes (PLWD) in the Ho Municipality, Ghana. The Open Public Health Journal. 2018;11(1).
Lokpo SY, Owiredu WK, Agordoh P, Agboli E, Amoo LNA, Noagbe M, et al. Cardio-Metabolic Risk Profile of a Diabetic Population in the Ho Municipality. Asian Journal of Research Reports in Endocrinology. 2018:1-11.
MOH. National Policy for the Prevention and Control of Chronic Non-Communicable Diseases in Ghana. 2012.
Makino M, Yoshimoto R, Ono M, Itoko T, Katsuki T, Koseki A, et al. Artificial intelligence predicts the progression of diabetic kidney disease using big data machine learning. Scientific reports. 2019;9(1):11862-.
Ho Teaching Hospital. Ho Teachning Hospital 2022 [Available from: https://www.hth.gov.gh/.
Reinhard H, Lajer M, Gall M-A, Tarnow L, Parving H-H, Rasmussen LM, et al. Osteoprotegerin and Mortality in Type 2 Diabetic Patients. Diabetes Care. 2010;33(12):2561-6.
Mulnier HE, Seaman HE, Raleigh VS, Soedamah-Muthu SS, Colhoun HM, Lawrenson RA. Mortality in people with Type 2 diabetes in the UK. 2006;23(5):516-21.
Owusu AY, Kushitor SB, Ofosu AA, Kushitor MK, Ayi A, Awoonor-Williams JK. Institutional mortality rate and cause of death at health facilities in Ghana between 2014 and 2018. PLoS One. 2021;16(9):e0256515.
Ang YG, Heng BH, Saxena N, Liew STA, Chong PN. Annual all-cause mortality rate for patients with diabetic kidney disease in Singapore. J Clin Transl Endocrinol. 2016;4:1-6.
Yoo H, Choo E, Lee S. Study of hospitalization and mortality in Korean diabetic patients using the diabetes complications severity index. BMC Endocrine Disorders. 2020;20(1):122.
González-Pérez A, Saez M, Vizcaya D, Lind M, Garcia Rodriguez L. Incidence and risk factors for mortality and end-stage renal disease in people with type 2 diabetes and diabetic kidney disease: a population-based cohort study in the UK. BMJ Open Diabetes Research &amp; Care. 2021;9(1):e002146.
Hsieh M-S, Hu S-Y, How C-K, Seak C-J, Hsieh VC-R, Lin J-W, et al. Hospital outcomes and cumulative burden from complications in type 2 diabetic sepsis patients: a cohort study using administrative and hospital-based databases. 2019;10:2042018819875406.
Magliano DJ, Harding JL, Cohen K, Huxley RR, Davis WA, Shaw JE. Excess Risk of Dying From Infectious Causes in Those With Type 1 and Type 2 Diabetes. Diabetes Care. 2015;38(7):1274-80.
Kvistholm JA, Nielsen EM, Björkman JT, Jensen T, Müller L, Persson S, et al. Whole-genome Sequencing Used to Investigate a Nationwide Outbreak of Listeriosis Caused by Ready-to-eat Delicatessen Meat, Denmark, 2014. Clin Infect Dis. 2016;63(1):64-70.
Afkarian M, Sachs MC, Kestenbaum B, Hirsch IB, Tuttle KR, Himmelfarb J, et al. Kidney disease and increased mortality risk in type 2 diabetes. J Am Soc Nephrol. 2013;24(2):302-8.
Lee S, Zhou J, Leung KSK, Wu WKK, Wong WT, Liu T, et al. Development of a predictive risk model for all-cause mortality in patients with diabetes in Hong Kong. BMJ Open Diabetes Research &amp; Care. 2021;9(1):e001950.
Wiyono S, Wibowo DS, Hidayatullah MF, Dairoh D. Comparative study of KNN, SVM and decision tree algorithm for student’s performance prediction. International Journal of Computing Science Applied Mathematics. 2020;6(2):50-3.
Yue C, Xin L, Kewen X, Chang S, editors. An intelligent diagnosis to type 2 diabetes based on QPSO algorithm and WLS-SVM. 2008 International Symposium on Intelligent Information Technology Application Workshops; 2008: IEEE.
Georga EI, Protopappas VC, Ardigo D, Marina M, Zavaroni I, Polyzos D, et al. Multivariate prediction of subcutaneous glucose concentration in type 1 diabetes patients based on support vector regression. IEEE journal of biomedical health informatics. 2012;17(1):71-81.

No competing interests reported.

Predictive models and determinants of mortality among T2DM patients in a tertiary hospital in Ghana, how do Machine Learning Techniques perform?

Status:

Version 1

Abstract

Figures

Introduction

Materials and methods

Study Design

Study area

Study population

Inclusion and Exclusion Criteria

Inclusion criteria

Exclusion criteria

Sample Size

Study variables

Data retrieval and management

Data Analysis

Ethical Consideration

Results

Discussion

Conclusion

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1