Development and verification of a machine learning-driven predictive model for cognitive impairment in Age-related Hearing Loss

doi:10.21203/rs.3.rs-4660830/v1

Download PDF

Research Article

Development and verification of a machine learning-driven predictive model for cognitive impairment in Age-related Hearing Loss

https://doi.org/10.21203/rs.3.rs-4660830/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Objective: The objective is to utilize machine learning algorithms to create a predictive model for cognitive impairment in age-relate hearing loss.

Methods:For this study, we gathered demographic information, conducted audiometric examinations, assessed cognitive abilities, and performed blood biochemical tests using data from NHANES. We then identified patients who fit the criteria based on the NHANES criteria. The LASSO regression method was employed to identify the determinants of cognitive impairment in age-related hearing loss. Additionally, five different machine learning algorithms were utilized to develop a predictive model for cognitive impairment in this population. Data from clinical trials were gathered between January 2024 and May 2024 to externally validate the model's dependability.

Results:The study comprised a total of 521 elderly adults with hearing loss, out of which 140 (26.8%) had cognitive impairment. The LASSO regression method was used to filter five factors: education level, alkaline phosphatase, globulin, creatinine, and eosinophil percentage. All of these predictors were then included in the machine learning approach for training the model.The XGB model exhibited consistent performance in both the training set (AUC=0.881) and the test set (AUC=0.868), while also demonstrating a moderate level of discrimination (AUC=0.706).

Conclusion:This study successfully developed a predictive model for cognitive impairment in age-related hearing loss using machine learning. The model demonstrates a satisfactory level of reliability and validity across several datasets. The utilization of machine learning models can serve as a potent instrument for healthcare providers in detecting cognitive impairment in the senior hearing loss population at an early stage.

age-related hearing loss

cognitive impairment

machine learning

predictive modelling

With the ongoing global aging trend, age-related hearing loss has emerged as a prevalent health issue among the elderly. According to the World Health Organization, the occurrence of age-related hearing loss rises rapidly as people get older. Globally, around 480 million individuals aged 65 years and above are affected by age-related hearing loss. Among those aged 70 years and above, the rate of hearing loss can be as high as 66 percent.[1, 2].

Research has demonstrated that geriatric individuals who are deaf have a notably greater occurrence of cognitive deficits in comparison to older adults who have normal hearing[3]. These cognitive impairments encompass memory loss, inattention, and reduction in executive functioning, among other effects. These impairments can significantly impact patients' quality of life and ability to carry out everyday activities. Elderly individuals with hearing loss have a distinct characteristic compared to other populations at risk for cognitive impairment, since their hearing loss can expedite the progression of cognitive decline[4, 5]. Each increment of 10 decibels in hearing loss is linked to a 1.3-fold rise in the likelihood of experiencing cognitive decline[6, 7]. Furthermore, older individuals who are deaf have a nearly 20% higher chance of having cognitive impairment compared to those with normal hearing, and this risk is more than doubled in individuals with severe hearing loss[8]. Furthermore, elderly individuals with hearing impairments may face challenges in accessing appropriate cognitive therapies, highlighting the necessity for more prompt and precise predictors and interventions[9].

The presence of comorbid cognitive impairment in elderly individuals who are deaf can result in several negative consequences, such as social isolation, mental health issues, and diminished capacity for self-care. These factors significantly impact their overall quality of life and ability to engage in social activities[10]. Furthermore, cognitive impairment might hasten the deterioration of mental faculties and raise the likelihood of illness in elderly individuals. Doctors can use predictive methods to identify older deaf patients who are at risk of cognitive deterioration. This early identification allows for prompt implementation of customised therapies and treatment programmes. Prompt intervention is crucial for mitigating cognitive deterioration and enhancing patients' overall well-being[11, 12].

Currently, the prediction model for cognitive decline in elderly deaf people is still in the first stage of investigation. Previous research primarily relies on conventional statistical techniques, but their capacity to make accurate predictions is constrained by criteria such as the amount of data available, the selection of relevant features, and the complexity of the models. Consequently, there are challenges related to low prediction accuracy and limited generalisation capability. Machine learning algorithms can effectively analyse extensive clinical data to identify probable patterns and traits, enabling precise prediction of the likelihood of cognitive impairment. The Nomogram, being a visual representation tool, may effectively present prediction outcomes and offer tailored intervention strategies for medical professionals and patients, thereby enhancing the applicability and efficacy of treatment. Hence, the utilisation of machine learning techniques to develop prediction models and nomograms for cognitive decline in elderly deaf individuals holds significant clinical importance and promising practical potential.

2.1 Study population and data

NHANES is a nationwide survey conducted in the United States to gather data from a representative sample of the population at a specific point in time. The National Center for Health Statistics of the Centers for Disease Control and Prevention (NCHS) provides access to it. The website for NCHS is https://www.cdc.gov/nchs/. The study protocol (protocol #2011-17; protocol #2011-17) received approval from the NCHS Research Ethics Review Board, and all participants gave written informed permission before taking part. The NHANES database was used to analyze data from audiometric examinations and cognitive assessments. Three cycles of data from 1999–2000, 2001–2001, and 2011–2012 were selected for the study. Data on demographics, cognitive assessment, audiometric screening, complete blood count screening, and blood biochemistry screening were collected.

The screening process for elderly patients with hearing loss involved the following criteria: age ≥ 60 years, completion of both audiometric examination and cognitive assessment. The patients' mean hearing thresholds were calculated, and cases with unilateral hearing loss as well as PTA, 6KHz, and 8KHz < 20dBHL were excluded based on the criteria outlined in the World Hearing Report 2001.

2.2 Study variables

This study included a total of 63 variables. These variables encompassed general data such as gender, age, education level, marital status, BMI, history of smoking, history of alcohol consumption, history of hypertension, and history of diabetes mellitus, totaling 9 variables. Additionally, examination data including pure tone audiometry, blood pressure, and 5 other variables were included. The audiometric data were used to evaluate the level of hearing loss in the patients. Paired t-tests were conducted on the hearing thresholds of the patients' right and left ears, and the mean value of the hearing thresholds for frequencies with differences was calculated. The average binaural hearing thresholds were calculated using frequencies that had no difference. Frequencies with differences were used to compare the hearing thresholds of the right and left ears. When the difference was more than 15 dBHL, the better ear was chosen. Conversely, the average of the right and left hearing thresholds was taken. The average hearing thresholds at frequencies of 0.5, 1, 2, and 4 KHz, as well as the severity of high-frequency hearing loss at frequencies of 4, 6, and 8 KHz, were determined. The mean systolic and diastolic blood pressures were calculated by averaging the values obtained from three measurements. The laboratory data consisted of complete blood counts and routine blood biochemistry, encompassing a total of 49 variables.

2.3 Definition of Cognitive Impairment

In this investigation, cognitive impairment was defined using the Digit Symbol Substitution Test (DSST), which is a component of the Wechsler Adult Intelligence Scale. The DSST assesses sustained attention and working memory[13]. The NHNAES database imposes a time constraint of 120 seconds for the test, and the score is determined by the total count of accurate matches during this time frame, with the maximum achievable score being 133. The assessment of low cognitive function was based on the 25th percentile cut-off point on the DSST exam, which aligns with the methods employed in previously published studies[14]. Individuals whose scores fell below the 25th percentile were classified as having low cognitive performance.The minimum score required to pass the DSST test was 33.

2.4 External validation

This study aimed to externally evaluate the performance of the model using clinical data obtained from aged deaf patients in three tertiary-level hospitals in Guangxi, China. The data was collected from January 2024 to May 2024. The purpose was to assess the model's capacity to generalize to other sample sets. Regarding cognitive assessment, we ensured consistency with the training phase of the model by employing the identical assessment procedure and cut-off criteria as the training dataset. The nadir criterion are outlined as follows:

①The inclusion criteria for this study are as follows: individuals must be at least 60 years old, have a type A tympanic chamber conductance map, exhibit binaural sensorineural hearing loss as confirmed by pure tone audiometry, and must sign an informed permission form to willingly participate in the study.

②Exclusion criteria for this study include individuals with combined brain injury, severe mental illness, severe speech disorders, sedation, inability to complete the necessary assessment or examination as required by the study, and non-age-related hearing loss conditions such as sudden deafness, noise deafness, pharmacological deafness, Meniere's disease, secretory otitis media, suppurative otitis media, otosclerosis, adhesive otitis media, sclerosis of the tympanic ventricle, and acoustic neuroma.

③Exclusion criteria: Participants with a percentage of missing data exceeding 50% will be excluded. Additionally, participants with missing critical data, such as hearing test results or cognitive assessment findings, will also be excluded.

Data from external sources were gathered in order to secure approval from the ethics committee for this study (approval number: KY2024-137), and all patients provided informed consent.

2.5 Statistical and model development

The study utilized SPSS 25.0 for statistical data description and R 4.3.2 software for machine learning modeling and analysis procedures. The initial data of the patients were analyzed by recording various statistics. For discrete variables, the number of values, missing values, and percentages were recorded. Continuous variables were described by calculating the maximum, minimum, median, and standard deviation, and the number of missing values was counted. Rank data and continuous variables that did not follow a normal distribution were analyzed using the Mann-Whitney U rank-sum test. Continuous variables that followed a normal distribution were analyzed using the two-independent-samples t-test. Discrete variables were analyzed using the chi-square test. The training and test sets were divided using stratified random sampling in a 7:3 ratio. Statistical descriptions of the two groups of patients were utilized to guarantee that the data sets were homogeneous. The predictors of cognitive impairment in age-related hearing loss patients were screened using one-way regression and LASSO regression. Prediction models were then constructed using five machine learning algorithms: logistics regression(Logit), random forests(RF), XGBoost(XGB), artificial neural networks(ANN), and support vector machines(SVM).

3.1 Baseline Characteristics of Patients

This study included a total of 3 cycles of data from 1999–2000, 2001–2002, and 2011–2012. After screening to exclude cases under 60 years of age and those with missing major indicators such as audiometric results and cognitive assessment, a total of 1294 cases were obtained. Further screening was done to exclude cases with no hearing loss, unilateral deafness, and outliers, resulting in a final inclusion of 521 patients. Among these patients, 332 (63.7%) were males and 189 (36.2%) were females. Additionally, 140 (26.8%) cases showed positive cognitive impairment.(Table 1 Baseline characteristics of patients)

Table 1

Baseline characteristics of patients
VariableNames	level	Overall(N = 521)	Cognitive Normal(N = 381)	Cognitive Impairment(N = 140)	p	SMD
			381	140
PTA		28 (20–83)	27 (20–83)	29 (20–68)	0.12	0.15
Age		65 (60–69)	64 (60–69)	65 (60–69)	0.41	0.08
BMI		28.25 (15.18–49.47)	28.42 (15.18–49.47)	28.05 (16.37–44.72)	0.61	0.05
SBP		131 (94–197)	129 (94–191)	134.5 (95–197)	< 0.01	0.27
DBP		72 (35–110)	72 (35–110)	71.5 (35–110)	0.9	0.01
Albumin		43 (26–50)	43 (35–50)	42 (26–50)	0.11	0.15
Alanine Aminotransferase		21 (8-191)	21 (8-191)	21 (9–70)	0.8	0.03
Aspartate Aminotransferase		23 (12–165)	23 (13–125)	23 (12–165)	0.05	0.17
Alkaline Phosphatase		71 (23–355)	70 (23–255)	74 (32–355)	< 0.01	0.34
Blood Urea Nitrogen		5 (1.43–34.99)	5 (1.43–17.14)	5.36 (1.79–34.99)	0.04	0.17
Ca		2.35 (1.9–2.65)	2.35 (1.9–2.65)	2.33 (2-2.55)	0.01	0.24
Total Cholesterol		5.17 (1.91–10.47)	5.22 (2.92–8.64)	5.16 (1.91–10.47)	0.31	0.1
Bicarbonate		25 (17–31)	25 (18–31)	24 (17–30)	0.09	0.17
Gamma Glutamyl Transferase		23 (6-556)	23 (7-394)	24.5 (6-556)	< 0.01	0.31
Blood Glucose		5.38 (2.33–34.25)	5.38 (3.11–34.25)	5.44 (2.33–29.2)	0.12	0.15
Fe		14.9 (3.9–47.6)	15 (3.9–47.6)	14.87 (4.48–45.85)	0.49	0.07
Lactate Dehydrogenase		133 (58–262)	132 (59–225)	138 (58–262)	< 0.01	0.3
P		1.16 (0.58–2.07)	1.16 (0.71–1.71)	1.13 (0.58–2.07)	0.13	0.14
Total Bilirubin		10.3 (3.4-61.56)	10.3 (3.4-46.17)	10.26 (3.4-61.56)	0.44	0.07
Total Protein		73 (59–91)	72 (59–90)	74 (61–91)	< 0.01	0.32
Triglycerides		1.55 (0.34–9.35)	1.55 (0.34–9.35)	1.47 (0.51–6.34)	0.16	0.14
Uric Acid		339 (89.2-618.6)	339 (166.5-618.6)	327.15 (89.2-594.8)	0.57	0.06
Creatinine		79.56 (35.4-840.68)	79.56 (35.4-480.01)	79.56 (35.4-840.68)	< 0.01	0.22
Na		139 (119-145.4)	139.3 (126-145.4)	139 (119-144.8)	0.43	0.07
K		4 (2.5–5.5)	4 (2.5–5.5)	4.08 (3.01–5.4)	0.27	0.11
Cl		103 (86–112)	103 (89–111)	103 (86–112)	0.13	0.14
Osmolality		280 (245–312)	280 (252–304)	280 (245–312)	0.38	0.08
Globulin		30 (18–51)	29 (18–45)	31 (21–51)	< 0.01	0.41
Glycohemoglobin		5.7 (4.4–16.9)	5.7 (4.4–16.9)	5.8 (4.5–13.4)	0.33	0.1
White Blood Cell Count		6.8 (3.2–34)	6.8 (3.3–34)	6.6 (3.2–17.5)	0.51	0.07
Lymphocyte Percent		28.7 (9.4–81.6)	28.7 (9.4–81.6)	29.15 (10-71.8)	0.15	0.14
Monocyte Percent		8 (0.8–20.6)	7.8 (1.9–17.5)	8.1 (0.8–20.6)	0.23	0.11
Neutrophils Percent		59.3 (13.1–87.2)	59.6 (13.1–87.2)	57.35 (22-84.3)	0.04	0.2
Eosinophils Percent		2.5 (0.2–17.5)	2.4 (0.3–13.1)	2.75 (0.2–17.5)	0.01	0.25
Basophils Percent		0.6 (0-8.5)	0.6 (0-8.5)	0.6 (0-3.1)	0.29	0.11
Lymphocyte Number		1.9 (0.7–27.7)	1.9 (0.7–27.7)	2 (0.7–6.7)	0.97	< 0.01
Monocyte Number		0.5 (0.1–1.6)	0.5 (0.2–1.6)	0.5 (0.1–1.4)	0.55	0.06
Neutrophils Number		4 (1.2–11.2)	4 (1.2–11.2)	3.8 (1.6–11.1)	0.19	0.13
Eosinophils Number		0.2 (0-1.1)	0.2 (0-1.1)	0.2 (0-1.1)	0.01	0.22
Basophils Number		0 (0-0.6)	0 (0-0.6)	0 (0-0.2)	0.09	0.18
Red Blood Cell Count		4.65 (2.78–6.31)	4.65 (2.99–6.31)	4.65 (2.78–5.78)	0.48	0.07
Hemoglobin		14.2 (9.6–18.3)	14.3 (9.6–18.3)	14.1 (9.6–18.3)	0.22	0.12
Hematocrit		42.1 (28.3–52.9)	42.1 (29.6–52.9)	41.75 (28.3–52.8)	0.46	0.07
Mean Cell Volume		91 (64.7-108.2)	90.9 (64.7-108.2)	91.9 (65.6-106.8)	0.73	0.03
Mean Cell Hemoglobin		31 (20.5–37.1)	31 (20.9–37.1)	31 (20.5–35.1)	0.68	0.04
Mean Cellular Hemoglobin Concentration		33.9 (29.2–36.9)	33.9 (29.2–36.9)	33.9 (31.2–35.9)	0.08	0.18
Red Cell Distribution Width		12.8 (11.4–19)	12.7 (11.4–17.3)	12.9 (11.4–19)	0.07	0.17
Platelet Count		234 (43–646)	234 (43–533)	230 (63–646)	0.71	0.04
Mean Platelet Volume		8.4 (6.4–13.5)	8.3 (6.4–13.5)	8.45 (6.6–13.1)	0.16	0.14
NLR		2.04 (0.16–9.25)	2.06 (0.16–9.25)	2 (0.3–8.25)	0.42	0.08
MLR		0.27 (0.06–1.29)	0.28 (0.06–0.67)	0.27 (0.06–1.29)	0.62	0.05
PLR		126.15 (5.13-592.22)	127.5 (5.13-592.22)	120.4 (19.7–440)	0.8	0.02
SII		472.5 (23.07-4023.75)	482 (23.07-4023.75)	436.26 (39.4–2904)	0.55	0.06
SIRI		1.06 (0.12–6.34)	1.08 (0.21–4.48)	0.99 (0.12–6.34)	0.6	0.05
Gender (%)	Male	332 (63.72)	232 (60.89)	100 (71.43)	0.03	0.22
	Female	189 (36.28)	149 (39.11)	40 (28.57)
Education (%)	Less Than 9th Grade	95 (18.23)	21 (5.51)	74 (52.86)	< 0.01	1.55
	9-11th Grade	96 (18.43)	63 (16.54)	33 (23.57)
	High School Grad	112 (21.50)	91 (23.88)	21 (15.00)
	Some College	128 (24.57)	119 (31.23)	9 (6.43)
	College Graduate or above	90 (17.27)	87 (22.83)	3 (2.14)
Marriage (%)	Married	347 (66.60)	265 (69.55)	82 (58.57)	0.04	0.32
	Widowed	43 (8.25)	30 (7.87)	13 (9.29)
	Divorced	68 (13.05)	50 (13.12)	18 (12.86)
	Separated	16 (3.07)	9 (2.36)	7 (5.00)
	Never married	31 (5.95)	16 (4.20)	15 (10.71)
	Living with partner	16 (3.07)	11 (2.89)	5 (3.57)
Hypertension (%)	No	249 (47.79)	184 (48.29)	65 (46.43)	0.78	0.04
	Yes	272 (52.21)	197 (51.71)	75 (53.57)
Diabetes (%)	No	387 (74.28)	292 (76.64)	95 (67.86)	0.05	0.2
	Yes	134 (25.72)	89 (23.36)	45 (32.14)
Smoking (%)	No	359 (68.91)	270 (70.87)	89 (63.57)	0.14	0.16
	Yes	162 (31.09)	111 (29.13)	51 (36.43)
Alcohol (%)	No	148 (28.41)	110 (28.87)	38 (27.14)	0.78	0.04
	Yes	373 (71.59)	271 (71.13)	102 (72.86)
Hearingloss Level (%)	20 ≤ PTA<35	386 (74.09)	287 (75.33)	99 (70.71)	0.65	0.15
	35 ≤ PTA<50	110 (21.11)	78 (20.47)	32 (22.86)
	50 ≤ PTA<65	22 (4.22)	14 (3.67)	8 (5.71)
	65 ≤ PTA<80	2 (0.38)	1 (0.26)	1 (0.71)
	80 ≤ PTA<95	1 (0.19)	1 (0.26)	0 (0.00)
HFHL level (%)	20 ≤ HFHL<35	70 (13.44)	46 (12.07)	24 (17.14)	0.58	0.2
	35 ≤ HFHL<50	150 (28.79)	114 (29.92)	36 (25.71)
	50 ≤ HFHL<65	163 (31.29)	117 (30.71)	46 (32.86)
	65 ≤ HFHL<80	108 (20.73)	80 (21.00)	28 (20.00)
	80 ≤ HFHL<95	22 (4.22)	17 (4.46)	5 (3.57)
	HFHL ≥ 95	8 (1.54)	7 (1.84)	1 (0.71)

3.2 Screening of model variables

The study employed one-way regression analysis to examine the factors linked to age-related hearing loss. These factors included gender, education, marriage, SBP, alkaline phosphatase, Ca, Gamma Glutamyl Transferase, Lactate Dehydrogenase, Total Protein, Creatinine, Globulin, Neutrophils percent, Eosinophils percent, Eosinophils number, and 15 additional factors. The LASSO regression was used to include predictors for screening, and model variables were selected using the one standard error approach (1se). The final model included five variables: Education, alkaline phosphatase, Globulin, Creatinine, and Eosinophils %. (Fig. 1: Variable selection using LASSO regression.(A) The LASSO coefficient curves are plotted for a total of 15 variables. Five variables with non-zero coefficients were chosen based on the best lambda values. (B) Cross-validation is used to determine the optimal lambda values for the LASSO model.)

(Fig. 1: Variable selection using LASSO regression.(A) The LASSO coefficient curves are plotted for a total of 15 variables. Five variables with non-zero coefficients were chosen based on the best lambda values. (B) Cross-validation is used to determine the optimal lambda values for the LASSO model.)

3.3 Model construction

This study assessed the utilization of five machine learning techniques for predicting cognitive impairment in elderly patients with hearing loss. The results indicated that the area under the curve (AUC) values for each model in the training set were as follows: RF AUC = 0.995, ANN AUC = 0.919, XGB AUC = 0.881, SVM AUC = 0.871, and Logit AUC = 0.866. In the testing set, the AUC values for each model were RF AUC = 0.856, ANN AUC = 0.877, XGB AUC = 0.868, SVM AUC = 0.852, and Logistics AUC = 0.859. The AUC values of the RF and ANN models in the test set exhibit more pronounced differences compared to the training set, indicating a potential issue of overfitting. On the other hand, the XGB, SVM, and Logit models show no significant reduction in performance in the test set, suggesting superior stability. (Fig. 2 displays the ROC of the five models in both the training set (A) and the test set (B). Table 2 presents the Confusion matrices of the five models in both the training set, test set and external validation set.)

(Fig. 2 displays the ROC of the five models in both the training set (A) and the test set (B). Table 2 presents the Confusion matrices of the five models in both the training set, test set and external validation set.)

Table 2

presents the Confusion matrices of the five models in both the training set, test set and external validation set
	Model	AUC	Sensitivity	Specificity	Pos Pred Value	Neg Pred Value	Precision	Recall	F1	Prevalence	Detection Rate	Detection Prevalence	Balanced Accuracy
Training Set	Logit	0.87	0.59	0.93	0.82	0.79	0.82	0.59	0.69	0.37	0.22	0.27	0.76
	RF	1	0.89	1	1	0.96	1	0.89	0.94	0.3	0.27	0.27	0.94
	XGB	0.88	0.6	0.93	0.82	0.8	0.82	0.6	0.69	0.37	0.22	0.27	0.76
	ANN	0.92	0.71	0.94	0.85	0.88	0.85	0.71	0.77	0.32	0.22	0.27	0.83
	SVM	0.87	0.55	0.93	0.84	0.76	0.84	0.55	0.67	0.4	0.22	0.27	0.74
Test Set	Logit	0.86	0.57	0.92	0.81	0.77	0.81	0.57	0.67	0.39	0.22	0.28	0.74
	RF	0.86	0.65	0.9	0.77	0.84	0.77	0.65	0.7	0.33	0.21	0.28	0.78
	XGB	0.87	0.57	0.92	0.81	0.77	0.81	0.57	0.67	0.39	0.22	0.28	0.74
	ANN	0.88	0.63	0.93	0.84	0.81	0.84	0.63	0.72	0.37	0.23	0.28	0.78
	SVM	0.85	0.56	0.91	0.81	0.75	0.81	0.56	0.66	0.4	0.22	0.28	0.73
Validation Set	Logit	0.69	0.68	0.71	0.43	0.87	0.43	0.68	0.53	0.25	0.17	0.38	0.69
	RF	0.69	0.52	0.78	0.74	0.57	0.74	0.52	0.61	0.55	0.28	0.38	0.65
	XGB	0.71	0.57	0.75	0.62	0.71	0.62	0.57	0.60	0.42	0.24	0.38	0.66
	ANN	0.68	0.67	0.72	0.48	0.85	0.48	0.67	0.56	0.28	0.19	0.38	0.70
	SVM	0.67	0.58	0.75	0.60	0.73	0.60	0.58	0.59	0.40	0.23	0.38	0.67

3.4 External validation results

The study collected data from 151 elderly patients with hearing loss to externally validate the models. Among these patients, 58 (38.4%) were found to have cognitive impairment. The results indicate that the AUC for the five machine learning models on the external validation set are as follows: RF AUC = 0.69, ANN AUC = 0.682, XGB AUC = 0.706, SVM AUC = 0.67, and Logit AUC = 0.692. Out of all the models, the XGB model demonstrates the highest performance and retains a moderate ability to distinguish in real data. (Fig. 3 displays the ROC for the five models on the external validation set)

The prevalence of age-related hearing loss has increased as the world population ages, resulting in a significant public health concern known as age-relate hearing loss. This decline in hearing not only impacts the overall well-being of patients, but it is also closely linked to the emergence of cognitive decline. Nevertheless, there remains a dearth of efficient prognostic instruments for promptly detecting age-related hearing loss in individuals with a heightened likelihood of cognitive decline in clinical settings. Traditional prediction models typically rely on clinical performance and lack objective data, resulting in limited predictive power and accuracy when dealing with large-scale, high-dimensional clinical data. In contrast, machine learning models are well-equipped to tackle these challenges. We extracted clinical objective data from the NHANES database using various machine learning algorithms. Subsequently, we employed one-way regression analysis to identify 15 indicators encompassing basic individual characteristics, lifestyle factors, and biochemical markers. We then applied LASSO regression to identify the factors that exhibited significant predictive value for cognitive impairment in elderly patients with hearing loss. A prediction model of cognitive impairment in elderly deaf people was constructed using five factors: education level, alkaline phosphatase, globulin, creatinine, and eosinophil percentage. Among the models evaluated, the XGB model consistently performs well in both the training and test sets. Although its performance decreases in the external validation set, the model still demonstrates more than moderate discriminative ability. This suggests that the model has reasonable predictive accuracy when applied to real clinical data.

The impact of cognitive reserve on cognitive impairment has been the focus of much research. Educational attainment, a significant indicator of cognitive reserve, has been demonstrated to have a substantial effect on cognitive impairment [15], which aligns with our study findings. There is a correlation between higher levels of education in older persons with hearing loss and a slower rate of cognitive decline. This correlation can be attributed to the positive impact of education on cognitive complexity and coping strategies. Hearing impairment has a negative impact on the ability of older individuals to engage in social activities and communicate effectively. This typically results in social challenges, which can further worsen cognitive decline[16]. Individuals with a greater degree of education are likely to have a larger social circle. This can be beneficial in maintaining a higher level of social engagement, even when experiencing hearing loss. Consequently, this can help reduce the adverse cognitive effects associated with hearing loss. When cognitive impairment arises, individuals with a higher level of education may be able to use their existing brain functions more effectively or discover alternative pathways for cognitive processing. This can lead to improved adaptation and resistance, ultimately enhancing an individual's ability to withstand cognitive decline[17]. For instance, these persons prioritize their health and are more inclined to actively pursue medical assistance and utilize assistive listening devices[18], such as hearing aids. Additionally, they are more inclined to pay closer attention to and comprehend information regarding hearing preservation. Consequently, they are more likely to adopt proactive measures at an earlier stage to mitigate the progression of hearing loss and its effects on cognitive function. Therefore, the level of education achieved may help mitigate the negative effects of hearing loss on cognitive functioning in older individuals. This is achieved by increasing cognitive reserve, improving social interactions, promoting healthy behaviors and psychological adjustments. Including this variable in the analysis enhances our understanding of the intricate connection between hearing loss and cognitive impairment.

Recent study has seen a growing number of scientists investigating the connection between liver and kidney function and cognitive decline. The liver and kidneys are vital metabolic organs in the body that are crucial for maintaining homeostasis and removing metabolic waste. If the function of these organs is compromised, it can have an impact on brain health and cognitive performance. Alkaline phosphatase and creatinine were identified as indicators of cognitive impairment in this investigation. An increased level of alkaline phosphatase in the body may indicate the existence of chronic illnesses, such as bone disease or impaired liver function, which might indirectly impact cognitive performance, particularly in older individuals. Research has indicated a potential correlation between bone health and cognitive performance[19]. For instance, an imbalance in the regulation of calcium metabolism and a lack of vitamin D may be connected to a deterioration in cognitive function, and both factors are indirectly linked to an increase in alkaline phosphatase activity[20]. Liver insufficiency can result in the buildup of toxins, metabolites, and various neurological disorders, which can impact brain function and hearing, thereby affecting cognitive performance[21, 22]. Similarly, hearing loss can also decrease the efficiency of nerve conduction, thereby affecting brain structure and function, including cognitive performance[23]. Creatinine levels serve as a crucial marker for evaluating the state of kidney function, and insufficient kidney function is regarded as a risk factor for hastened cognitive decline [24]. Renal insufficiency is linked to elevated levels of inflammation in the body, which can have a negative effect on vascular health by encouraging the development of atherosclerosis [25], including in the blood arteries of the brain. Hearing loss has been linked to impaired intracranial vascular condition, which can impact the circulation of blood to the ears and brain, consequently affecting both hearing and cognitive abilities[26]. Maintaining proper electrolyte balance and acid-base balance in the body is crucial for normal nerve conduction and brain function[27]. Elevated levels of creatinine may indicate a disruption in these balances, potentially impacting the function of auditory nerves and cognitive areas of the brain. Creatinine and alkaline phosphatase likely play a role in the connection between hearing loss and cognitive impairment through various physiological and pathological routes. Gaining a comprehensive understanding of these mechanisms would not only clarify the biological foundation of the cognitive impacts of hearing loss, but also offer crucial insights for the creation of specific preventative and treatment approaches. Future research will focus on thoroughly investigating the significance of these indicators in order to gain a comprehensive understanding of and effectively manage cognitive decline in older individuals.

Chronic inflammation has been established to be one of the key risk factors for cognitive impairment, which is consistent with the results of this study that found a link between globulin and eosinophil percentages and cognitive impairment. Both may play a bridge function between hearing loss and cognitive impairment. Changes in globulin and eosinophil percentage can reflect the body's inflammatory state or immune function, and their elevation usually reflects the presence of chronic inflammation or increased immune activity in the body, alterations that are thought to correlate with hearing loss and cognitive impairment[28]. Hearing loss itself is connected with increased levels of systemic inflammation, and inflammatory states may impact the health of auditory circuits. For example, otitis media can aggravate hearing loss, which can disrupt hearing-related cognitive processing, such as language comprehension and speech processing [29, 30]. On the other side, social isolation and communication difficulties due to hearing loss may induce a chronic stress reaction, which in turn increases the body's inflammatory response and impacts brain function, including memory and learning ability[31]. A chronic inflammatory condition will impair the integrity of the blood-brain barrier, allowing substances that would not usually enter the brain, such as inflammatory mediators, to enter the brain and impact the function and survival of nerve cells[32]. Globulin and eosinophil percentages as inflammatory indicators may reflect this altered barrier function. In addition, hearing loss may make it essential for older persons to commit more cognitive resources to interpreting language in everyday communication, which would raise their cognitive load[33]. Prolonged high cognitive load leads to overconsumption of cognitive resources and accelerates cognitive decline, in a process in which the body's immunological regulation and inflammatory response may further impact the management and allocation of cognitive resources[34]. This finding of peripheral inflammatory indicators such as globulins and eosinophils as predictors could help identify underlying inflammatory or immune-mediated pathological processes that may be co-factors of hearing loss and cognitive decline, providing clues for the development of more effective treatment and management strategies.

To summarize, the importance of five indicators - educational level, alkaline phosphatase, creatinine, eosinophil percentage, and globulin - offers us fresh insights into potential connections between cognitive reserve, hepatic and renal abnormalities, and inflammatory and immune status with cognitive impairment in elderly individuals with hearing loss. This discovery offers valuable insights for further investigating the causes of the disease and developing strategies to enhance our understanding and avoid the onset of cognitive decline in elderly individuals with hearing loss.

Limitation

The study yields valuable insights, although it is important to understand certain limitations. Initially, our investigation employed a cross-sectional design, which precluded us from establishing causation. Furthermore, the data used for training the model were acquired retrospectively from a database, which had a slight impact on the diversity of variables. Future studies should utilize large-sample, multicenter prospective cohorts and conduct a full assessment of cognitive impairment to improve the prediction of cognitive impairment in the elderly population with hearing loss.

The educational level, alkaline phosphatase, globulin, creatinine, and eosinophil percentage demonstrated significant predictive value for cognitive impairment in elderly individuals with hearing loss. Out of the models created in this study, the XGB model demonstrated the highest performance in the external validation set and exhibited a modest level of predictive ability.

Ethical Statement and Inform Consent

This study received ethical approval from Ruikang Hospital Affiliated to Guangxi University Of Chinese Medicine IRB (Approval No. KY2024-137) . Patient data will not be shared with third parties.

Publication Consent

All authors have read and agreed to the publication of this manuscript.

Conflict of Interests

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

Data Availability Statement

All data generated during this study are included in this article. Further enquiries can be directed to the corresponding author.

Founding

This research was supported by two funds from the Guangxi University of Chinese Medicine. The first support came from the 2023 Doctoral Research Startup Fund for the project titled "Development of a Predictive Model for Mild Cognitive Impairment in Age-Related Hearing Loss Using Machine Learning," under the project number 2023BS056. Additionally, our project has also received support from the High-Level Talent Cultivation Innovation Team Fund, under grant number 2022A010.

Author`s Contributions

RL,JS, and YS initiated the project and designed the research framework. YZ , HL , SQ , WW , and CW were responsible for collecting and entering data for external validation. HT took charge of data verification. ZZ handled the construction of the model, data analysis, and the drafting of the manuscript. RL and JS participated in the review and revision of the manuscript and supervised all aspects of the work. YS and ZZ contributed equally to this study.

Acknowledgements

We express our heartfelt gratitude to all the clinical workers involved in the data collection process, whose diligent work made this research possible. Special thanks to Dr. Shuai Jin for his expert guidance in the methodology of machine learning techniques, whose insightful and valuable advice greatly enriched the content and methods of our study. We deeply appreciate the support and assistance from all of them.

Goman AM, Lin FR. Prevalence of Hearing Loss by Severity in the United States. Am J Public Health. 2016;106(10):1820–2. https://doi.org/10.2105/AJPH.2016.303299.
Rutherford BR, Brewster K, Golub JS, Kim AH, Roose SP. Sensation and Psychiatry: Linking Age-Related Hearing Loss to Late-Life Depression and Cognitive Decline. Am J Psychiatry. 2018;175(3):215–24. https://doi.org/10.1176/appi.ajp.2017.17040423.
Loughrey DG, Kelly ME, Kelley GA, Brennan S, Lawlor BA. Association of Age-Related Hearing Loss With Cognitive Function, Cognitive Impairment, and Dementia: A Systematic Review and Meta-analysis. JAMA Otolaryngol Head Neck Surg. 2018;144(2):115–26. https://doi.org/10.1001/jamaoto.2017.2513.
Croll PH, Vinke EJ, Armstrong NM, et al. Hearing loss and cognitive decline in the general population: a prospective cohort study. J Neurol. 2021;268(3):860–71. https://doi.org/10.1007/s00415-020-10208-8.
Liberman MC, Kujawa SG. Cochlear synaptopathy in acquired sensorineural hearing loss: Manifestations and mechanisms. Hear Res. 2017;349. https://doi.org/10.1016/j.heares.2017.01.003. :138 – 47.
Torrente MC, Vergara R, Moreno-Gómez FN, et al. Speech Perception and Dichotic Listening Are Associated With Hearing Thresholds and Cognition, Respectively, in Unaided Presbycusis. Front Aging Neurosci. 2022;14:786330. https://doi.org/10.3389/fnagi.2022.786330.
Deal JA, Betz J, Yaffe K, et al. Hearing Impairment and Incident Dementia and Cognitive Decline in Older Adults: The Health ABC Study. J Gerontol Biol Sci Med Sci. 2017;72(5):703–9. https://doi.org/10.1093/gerona/glw069.
Yuan J, Sun Y, Sang S, Pham JH, Kong WJ. The risk of cognitive impairment associated with hearing function in older adults: a pooled analysis of data from eleven studies. Sci Rep. 2018;8(1):2137. https://doi.org/10.1038/s41598-018-20496-w.
International AD. World Alzheimer Report 2023 Reducing dementia risk: never too early, never too late,2023.
Atef RZ, Michalowsky B, Raedke A, et al. Impact of Hearing Aids on Progression of Cognitive Decline, Depression, and Quality of Life Among People with Cognitive Impairment and Dementia. J Alzheimers Dis. 2023;92(2):629–38. https://doi.org/10.3233/JAD-220938.
Liao YY, Tseng HY, Lin YJ, Wang CJ, Hsu WC. Using virtual reality-based training to improve cognitive function, instrumental activities of daily living and neural efficiency in older adults with mild cognitive impairment. Eur J Phys Rehabil Med. 2020;56(1):47–57. https://doi.org/10.23736/S1973-9087.19.05899-4.
Livingston G, Huntley J, Sommerlad A, et al. Dementia prevention, intervention, and care: 2020 report of the Lancet Commission. Lancet. 2020;396(10248):413–46. https://doi.org/10.1016/S0140-6736(20)30367-6.
Jasinski LJ, Berry DT, Shandera AL, Clark JA. Use of the Wechsler Adult Intelligence Scale Digit Span subtest for malingering detection: a meta-analytic review. J Clin Exp Neuropsychol. 2011;33(3):300–14. https://doi.org/10.1080/13803395.2010.516743.
Dong X, Li S, Chen J, Li Y, Wu Y, Zhang D. Association of dietary ω-3 and ω-6 fatty acids intake with cognitive performance in older adults: National Health and nutrition examination Survey (NHANES) 2011–2014. Nutr J. 2020;19(1):25. https://doi.org/10.1186/s12937-020-00547-7.
Gu H, Yao X, Diao C, et al. Global cognitive function is associated with sex, educational level, occupation type, and speech recognition rate in older Chinese adults: a single-center, prospective, cross-sectional study. BMC Geriatr. 2022;22(1):947. https://doi.org/10.1186/s12877-022-03603-5.
Bowl MR, Dawson SJ. Age-Related Hearing Loss. Cold Spring Harb Perspect Med. 2019;9(8):a033217. https://doi.org/10.1101/cshperspect.a033217.
Alattar AA, Bergstrom J, Laughlin GA, et al. Hearing Impairment and Cognitive Decline in Older, Community-Dwelling Adults. J Gerontol Biol Sci Med Sci. 2020;75(3):567–73. https://doi.org/10.1093/gerona/glz035.
Weycker JM, Dillard LK, Pinto A, Fischer ME, Cruickshanks KJ, Tweed TS. Factors Affecting Hearing Aid Adoption by Adults With High-Frequency Hearing Loss: The Beaver Dam Offspring Study. Am J Audiol. 2021;30(4):1067–75. https://doi.org/10.1044/2021_AJA-21-00050.
Trivedi T, Guise TA. Systemic effects of abnormal bone resorption on muscle, metabolism, and cognition. Bone. 2022;154:116245. https://doi.org/10.1016/j.bone.2021.116245.
Beauchet O, Launay CP, Galery K, et al. Effects of Vitamin D and Calcium Fortified Yogurts on Gait, Cognitive Performances, and Serum 25-Hydroxyvitamin D Concentrations in Older Community-Dwelling Females: Results from the GAit, MEmory, Dietary and Vitamin D (GAME-D2) Randomized Controlled Trial. Nutrients. 2019;11(12):2880. https://doi.org/10.3390/nu11122880.
Xia L, Wang T, Chen X, et al. Serum Prealbumin as a Potential Predictive Factor for Age-Related Hearing Loss: A Retrospective Study. Ear Nose Throat J. 2024;1455613241254241. https://doi.org/10.1177/01455613241254241.
Reja M, Phelan LP, Senatore F, Rustgi VK. Social Impact of Hepatic Encephalopathy. Clin Liver Dis. 2020;24(2):291–301. https://doi.org/10.1016/j.cld.2020.01.008.
Slade K, Plack CJ, Nuttall HE. The Effects of Age-Related Hearing Loss on the Brain and Cognitive Function. Trends Neurosci. 2020;43(10):810–21. https://doi.org/10.1016/j.tins.2020.07.005.
Xiao Y, Devakumar V, Xu L, Liu L, Mo H, Hong X. Elevated serum creatinine levels and risk of cognitive impairment in older adults with diabetes: a NHANES study from 2011–2014. Front Endocrinol (Lausanne). 2023;14:1149084. https://doi.org/10.3389/fendo.2023.1149084.
Düsing P, Zietzer A, Goody PR, et al. Vascular pathologies in chronic kidney disease: pathophysiological mechanisms and novel therapeutic approaches. J Mol Med (Berl). 2021;99(3):335–48. https://doi.org/10.1007/s00109-021-02037-7.
Parker TD, Hardy C, Keuss S et al. Peripheral hearing loss at age 70 predicts brain atrophy and associated cognitive change. J Neurol Neurosurg Psychiatry. 2024:jnnp-2023-333101 [pii]. https://doi.org/10.1136/jnnp-2023-333101.
Rosner MH, Husain-Syed F, Reis T, Ronco C, Vanholder R. Uremic encephalopathy. Kidney Int. 2022;101(2):227–. https://doi.org/10.1016/j.kint.2021.09.025. 41.
Yang H, Liao Z, Zhou Y, Gao Z, Mao Y. Non-linear relationship of serum albumin-to-globulin ratio and cognitive function in American older people: a cross-sectional national health and nutrition examination survey 2011–2014 (NHANES) study. Front Public Health. 2024;12:1375379. https://doi.org/10.3389/fpubh.2024.1375379.
Zhou T, Mao J, Zhu P, Yu X, Yang X. Association between the systemic immuno-inflammation index and hearing loss: result from NHANES 2009–2018. Front Neurol. 2024;15:1369492. https://doi.org/10.3389/fneur.2024.1369492.
Watson N, Ding B, Zhu X, Frisina RD. Chronic inflammation - inflammaging - in the ageing cochlea: A novel target for future presbycusis therapy. Ageing Res Rev. 2017;40:142–8. https://doi.org/10.1016/j.arr.2017.10.002.
García-García I, Donica O, Cohen AA, et al. Maintaining brain health across the lifespan. Neurosci Biobehav Rev. 2023;153:105365. https://doi.org/10.1016/j.neubiorev.2023.105365.
Garcia-Escobar G, Manero RM, Fernández-Lebrero A, et al. Blood Biomarkers of Alzheimer's Disease and Cognition: A Literature Review. Biomolecules. 2024;14(1):93. https://doi.org/10.3390/biom14010093.
Yu RC, Proctor D, Soni J, et al. Adult-onset hearing loss and incident cognitive impairment and dementia - A systematic review and meta-analysis of cohort studies. Ageing Res Rev. 2024;98:102346. https://doi.org/10.1016/j.arr.2024.102346.
Geng C, Chen C. Association between elevated systemic inflammatory markers and the risk of cognitive decline progression: a longitudinal study. Neurol Sci. 2024. https://doi.org/10.1007/s10072-024-07654-x.

No competing interests reported.

Download PDF

Editorial decision: Revision requested
12 Jul, 2024
Editor assigned by journal
11 Jul, 2024
Submission checks completed at journal
11 Jul, 2024
First submitted to journal
29 Jun, 2024

You are reading this latest preprint version

Development and verification of a machine learning-driven predictive model for cognitive impairment in Age-related Hearing Loss

Status:

Version 1

Abstract

Figures

Introduction

Methods

2.1 Study population and data

2.2 Study variables

2.3 Definition of Cognitive Impairment

2.4 External validation

2.5 Statistical and model development

Result

3.1 Baseline Characteristics of Patients

3.2 Screening of model variables

3.3 Model construction

3.4 External validation results

Discussion

Limitation

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1