Applying Artificial Intelligence to Predict Self-Reported Poor Health among Black and Hispanic Caregivers with Mild Cognitive Impairment

Abstract

We applied artificial intelligence techniques to build correlate models that predict general poor health in a national sample of caregivers with mild cognitive impairment (MCI). Our application of deep learning identified age, duration of caregiving, amount of alcohol intake, weight, myocardial infarction (MI) and frequency of MCI symptoms for Blacks and Hispanics whereas frequency of MCI symptoms, income, weight, coronary heart disease (CHD), age, and use of e-cigarette for the others as the strongest correlates of poor health among 81 variables entered. The application of artificial intelligence efficiently provided intervention strategies for Black and Hispanic caregivers with MCI.

1. Introduction

Approximately one in five Americans age 65 or older have preclinical Alzheimer’s disease (AD), mild cognitive impairment (MCI) [1]. The prevalence of MCI is higher in Hispanics and Blacks than non-Hispanic Whites (27.3% vs. 19.1%) [2]. Two proteins including β-amyloid (plaques) and tau (tangles) abnormally increase in number during aging and in AD; no medications are currently available for MCI [3].

Deep learning (a subset of machine learning where artificial neural networks, algorithms inspired by the human brain, learn from large amounts of data) has the potential to assist personalized recommendations based on their contextual and lifestyles and situational factors among caregivers of AD. Deep learning, representing the latest iteration in a progression of artificial intelligence (AI) technologies, has made advances in solving problems which have resisted the best attempts of the AI community [4]. Despite its popularity, deep learning has rarely been applied in behavioral medicine. The purpose of the study was to build deep learning algorithms to predict self-reported poor health based on aging, behavioral, and demographic factors as the groundwork for future targeted interventions for Black and Hispanic caregivers with MCI in the U.S.

2. Methods

We applied a data mining process to a publicly available national dataset, a survey of 1,373,755 Americans from the Behavioral Risk Factor Surveillance System (BRFSS) from 2016 to 2018, to build a model for predicting self-reported poor health. BRFSS is a U.S. phone based survey to collect data regarding health-related risk behaviors. We used R 3.6.1, Weka 3.8.3, and Deep Neural Network-BigML to conduct AI based analytics. We extracted 275 demographic, behavioral and physio-psycho social variables using R 3.6.1. First, 81 of total 275 variables were selected by caregiving and behavior experts as relevant to poor health based on the literature [1,3,5,6]. During feature selection, we applied deep learning in BigML, which evaluates the importance of attributes by predictive ability of each feature considering redundancy between them.

The outcome variable was the dichotomized self-reported poor health [the respondents were asked to rank their general health on a five-point Likert scale (1 excellent, 2 very good, 3 good, 4 fair, 5 poor), 4 or 5 corresponded to a “Yes”]. The respondents answered “Yes” to the question [During the past 12 months, have you experienced confusion or memory loss that is happening more often or is getting worse?] were identified as caregivers with MCI. BigML with default configuration were chosen. We applied correlation, entropy, and learner based techniques to create different views of the dataset to check algorithms dependency and selected eight final variables based on the criteria of clinical meaningfulness based on an understanding of the literature on reasons for poor health among people with MCI [1,3,5,6]. Next, we organized the variables into three categories: demographic, physiological, and behavioral factors.

Next, we iteratively applied C4.5 algorithms to build the prediction models for self-reported poor health among caregivers with MCI. For cross-validation (10-fold), the dataset was randomly divided into training and evaluating datasets for the model validation before applying the algorithms. We chose the final models based on the model predictive accuracy, interpretability, and clinical meaningfulness, and the area under the receiver operating characteristic curve (AUC). Lastly, neuro-science nursing and behavioral science experts interpreted the models according to clinical meaningfulness.

3. Results

Study participants (n=897, 0.07%) were identified as caregivers with MCI. Age 65 or older were 44.26% (n=397). Over half of caregivers with MCI were female (58.42%, m=523). One of ten participants (n=77, 8.58%) identified themselves as Black of Hispanic caregivers. Half of caregivers with MCI reported their health as poor (n=438, 48.83%). Descriptive statistics for the study variables are summarized in Table 1.

Table 1.

Descriptive Statistics for Study Participants (n=897)

		Blacks or Hispanics	Others†	χ2 test
		n=77 (%)	n=820 (%)	p-value
MCI symptoms	Always, Usually	15 (19.48)	116 (14.15)	0.06
	Sometimes	22 (28.57)	164 (20.00)
	Rarely, Never	40 (51.95)	529 (65.51)
Income	<25K	39 (50.65)	282 (34.39)	0.00
	≥25K	27 (35.06)	429 (52.32)
Exercise	Yes	43 (55.84)	546 (66.59)	0.07
E-Cigarette use	Yes	1 (1.30)	43 (5.24)	0.21
Poor health	Yes	42 (54.55)	396 (48.29)	0.35

^†Others: non-Hispanic whites, Asian or Pacific Islander, or Aleutian/Eskimo/American Indian

While age (importance 12.02% for Blacks and Hispanics) and weight (importance 8.96% for Blacks and Hispanics, 8.55% for others) also selected by deep learning algorithms, they were excluded in the final prediction model due to their relative lack of interpretability (e.g., proxies for number of chronic illnesses) and applicability.

4. Discussion

Among the many risk factors for self-reported poor health, deep learning found that frequency of MCI symptoms interfering with social activities was a key determinant (variable importance: 6.37% for Blacks and Hispanics, 11.17% for others, Figure 1) among US caregivers with MCI. Consistent with the literature, our findings indicate that frequency of MCI symptoms is a strong predictor of functional ability and the need for care in older people [5]. MCI is also associated with impaired activities of daily living, unplanned hospital admissions, and the need for long-term care [5]. Our study found that among Black and Hispanic caregivers with MCI under poverty (annual household income ≤$25,000), whose MCI symptoms have interfered with social activities during the past 12 months, were likely to report being in poor health regardless of their caregiving duration whether inexperienced or experienced (figure 2).

Figure 1.

Ranking of Factor Importance for Poor Health Applying Deep Learning

Figure 2.

Prediction Models for Self-Reported Poor Health for Caregivers with MCI

Novel findings were observed for those who used electronic cigarettes. Among caregivers who exercised and experienced MCI symptoms interfering with social activities, the use of e-cigarettes increased their probability of experiencing poor health (Figure 2). Although electronic cigarettes have not been on the market long, there is clear evidence that the aerosols of most electronic nicotine delivery systems contain toxic chemicals [7]. Considering the paths among these predictors will be an important next step. Is exercise enough to improve the general health for caregivers with MCI? If not, such a pattern would suggest behavioral interventions for cessation of e-cigarette smoking as treatments when MIC symptoms occur, rather than only exercise counseling. Targeting patients whose social activities are curtailed due to their MCI symptoms may be appropriate for effective interventions. Additionally, this study is the first of its kind to document the prevalence and impact of MCI among caregivers. MCI can affect a person’s ability to learn new skills and complete tasks [6]. Understanding the prevalence, frequency, and impact of MCI symptoms can offer insight into the types of interventions that can support needs and explain why some caregivers are more burdened than others.

We used the Deep Neural Network function of the BigML platform to automate the discovery of the parameterization for our classification task. This approach obviates a primary practical challenge to using deep learning: determining the optimal settings for the hyper-parameters of the model (e.g., the appropriate number of hidden layers) [4]. The lack of transparency and interpretability of the model’s predictions is a major limitation of such methods; nonetheless, the combined use of traditional machine learning algorithms such as C4.5 with the Deep Neural Network automatic optimization features of emerging platforms like BigML is suggested for similar studies.

5. Conclusion

Our deep learning algorithms revealed the strongest predictors of self-reported poor health among caregivers with MCI to be aging, frequency of MCI symptoms interference with social activities, exercise, e-cigarette use, and income. This new knowledge adds insights about behavior and supports symptom science interventions for caregivers with MCI and other aging-related future interventions.