Early Alert of Elderly Cognitive Impairment using Temporal Streaming Clustering

Abstract

more than 44 million people have been diagnosed with dementia worldwide, and this number is estimated to triple by next three decades. Given this increasing trend of older adults with cognitive impairment (CI; dementia and mild cognitive impairment) and its significant underdiagnosis, early identification of CI and understanding its progression is a critical step towards a better quality of life for the aging population. Early alert of individual health changes could facilitate better ways for clinicians to diagnose CI in its early stages and come up with more effective treatment plans. However, there is a lack of approaches to characterize patient health conditions accounting for temporal information in an unsupervised manner. Limited CI cases and its costly ascertainment in clinical settings also make unsupervised learning more promising in CI research. In this paper, a streaming clustering model was used to determine distinct patterns of older adults’ health changes from their clinical visits in Mayo Clinic Study of Aging. The streaming clustering was also examined to study its ability to generate early alerts for potential incidents of CI. Our analysis demonstrated that temporal characteristics incorporated in a streaming clustering model has a promising potential to increase power in predicting CI.

I. Introduction

Cognitive impairment (CI, here in defined as mild cognitive impairment [MCI] and dementia) is one of the most predominant health problems in the aging population. The most common type of dementia is Alzheimer’s disease (AD) which currently affects more than 6 million people in the United States and will affect an estimated 15 million by 2060 [1]. These large numbers have resulted in a tremendous burden on patients, families, and the health system, with an annual estimate of 18.5 billion hours of unpaid care at a value of $234 billion [2]. Delays in clinical diagnosis of a patient with CI results in missed opportunities for appropriate care plans, leading to adverse clinical outcomes. Even though many people are affected by AD and related dementias, more than 40% of patients with CI are undiagnosed [3]. In addition, clinical diagnosis usually appears late in the process of cognitive decline or after the disease is severely progressed, and the opportunity to maximize the quality of life is no longer achievable [4–6]. Early detection of cognitive decline will give clinicians more time to develop better treatment plans for CI patients and lead to a better quality of life for the aging population [7].

The complexity of CI diagnosis results in missed opportunity for early diagnosis and care plans that can maximize CI patients’ quality of life. Our preliminary findings revealed that early signs of CI are available in electronic health records (EHRs) several years before the clinical diagnosis of CI [8,9]. A promising approach to predict CI is to study longitudinal trajectories of different clinical assessments of CI and cognitively unimpaired (CU) patients. Several studies have shown that trajectory analysis may detect early signs of CI much earlier than the actual dementia diagnosis [10,11]. In addition, one of our studies on activities of daily living demonstrated distinct temporal groups between CU and CI older adults multiple years before clinical diagnosis [8, 9].

Recent advances in machine learning models have demonstrated capabilities in analyzing EHR data [12–16]. EHR data is a rich source of information that can be harnessed to provide earlier detection and identification of cognitive declines in older patients. However, there is a lack of systematic analysis of patient data related to CI in routinely collected EHRs, and how their temporal patterns are associated with the development of CI. Although predicting CI and understanding the progression from CU to MCI and dementia utilizing EHRs is a challenging and largely unexplored task, we believe temporal models such as streaming clustering can be very effective in capturing temporal changes in patient data to predict early stages of CI. This study systematically investigated the application of EHR data analyses and streaming clustering models for clustering and predicting CU, MCI, and dementia patients.

Existing models for risk prediction have mainly focused on static models, which do not account for temporal information of health status [17–19]. Despite some early success in modeling patients’ future risk, the informative longitudinal patterns and dynamic changes in patient’s health status were not captured. Most recent studies on temporal-based models on MCI and dementia progression are based on supervised learning. The majority of the focus is on recurrent neural networks (RNN) and their variants taking advantage of their ability to learn complex nonlinear relationships and sequential patterns [20].

Supervised learning models require human effort to label the data which increases the cost and complexity of the process. Unsupervised learning models, in contrast, work on unlabeled data to discover structures and relationships between clusters. In addition, classification algorithms require long time for training and validation whereas unsupervised techniques can be run in real time without training phase. Unsupervised learning models such as clustering have been used to look for groups in dementia and MCI patients [21–24]. Traditional clustering algorithms do not account for temporal information because the model is static and does not adapt over time as the data evolve in the feature space. Streaming Clustering algorithms are incremental models that search for groups and trends in the data streams. They have the advantage of taking the temporal information into account as well as adapting with the data over time. Trajectories and trends can be analyzed using streaming clustering, which is not feasible in the traditional clustering techniques. Trajectories can be analyzed in a streaming clustering setting by observing the cluster change over time. For example, in dementia and MCI detection, if a patient’s health data start at a particular cluster and change cluster for one of the future visits, that would trigger an alarm to investigate the reason for the change. This change could be related to early cognitive decline and can be used to flag an alarm that indicates MCI or dementia.

Temporal-based unsupervised techniques such as streaming clustering have not been studied in the context of CI progression and prediction. Streaming clustering algorithms demonstrated their capabilities in analyzing temporal data and predicting early signs of illness in eldercare technology [25, 26]. Streaming clustering models have been applied on datasets where each subject is represented with a single data sample (represent each patient with one visit) or all data samples are from the same subject (all visits from the same patient) [25, 26]. The challenge in this study is that each subject (patient) has multiple health visits. Thus, combining the visits of a patient with a single representation could result in losing some of the temporal characteristics. We also do not have enough health visits to develop a model for each patient. Therefore, we adapted a streaming clustering model incorporating temporal patient health status to identify distinct groups in CU and CI patients as well as analyze patient characteristics. Each patient has multiple health visits and the number of visits varies among patients. We also investigated the potential of streaming clustering to predict early signs of MCI and dementia. The model is tested on a population-based cohort with a comprehensive periodic cognitive assessment of MCI and dementia. We also proposed a heuristic approach to monitor the cluster change of the streaming clustering algorithm and predict early signs of MCI and dementia. We compared our model performance with other supervised learning models and achieved higher accuracy and F1 score.

II. Materials and Methods

In this section, we first describe the dataset used in our model. Then, we describe the streaming clustering algorithm used in this study. This study includes two main components: patient clustering, and dementia and MCI prediction. In both clustering and prediction, we adapted the MU Streaming Clustering (MUSC) algorithm [25] described in section II.C. We considered a cluster change between visits for a patient as a sign of change in the cognitive status.

A. Data

This study was approved by the Mayo Clinic Institutional Review Board and the Olmsted Medical Center Institutional Review Boards. We used data from the Mayo Clinic Study of Aging (MCSA) [27]. The MCSA is a prospective population-based cohort study with comprehensive periodic cognitive assessments (at baseline and repeated every 15 months), initiated in 2004 to investigate the epidemiology of MCI. Eligible persons from the Olmsted County, Minnesota population, were randomly selected and evaluated comprehensively in person by three independent evaluators, using the clinical dementia rating scale, a neurological evaluation, and neuropsychological testing. A consensus committee used previously published criteria to diagnose the participants with CU, MCI, or dementia. MCI is diagnosed according to published criteria [28], and dementia is diagnosed according to DSM-IV criteria [29]. In addition to cognitive assessment, other data elements are abstracted from the medical records or through interviews and questionnaires (e.g., education, body mass index, comorbid conditions, neuropsychiatric symptoms, vascular risk factors). The MCSA cohort comprises 6,185 unique patients with 26,807 visits (4.3 average visits per patient). Among these, 3,070 patients were female (49.6%), and 729 patients (11.6%) had progressed to dementia. The median age of the cohort at baseline was 73. Table 1 contains the basic characteristics of the MCSA cohort used as input variables in the study.

Table 1.

Input Variables

Patient Demographics

Age, Sex, Race, Ethnicity

Physical characteristics

BMI, Smoking status, Alcohol problem

Sleep apnea, Hypertension, Dyslipidemia, Atrial fibrillation, Angina, Congestive heart failure, Coronary artery disease, Myocardial infarction, Coronary artery bypass graft, Diabetes, ESSscore

Neuropsychiatric characteristics

Delusions, Hallucinations, Agitation, Depression, Anxiety, Euphoria, Apathy, Disinhibition, Irritability/lability, Motor behavior, Nighttime behavior, Appetite/eating change, BDI-II grand total, BDI depression (Total >=13), BAI total (0–63)

Social characteristics

Education, Occupation, Marital status, Personal care (as extracted from the clinical dementia rating scale)

Functional status

FAQ Total Score (0–30), ECog-12

Abbreviations: ESSscore: Epworth Sleepiness Scale score (0–24), BDI: Beck Depression Inventory scores, BAI: Beck Anxiety Inventory, Functional Activities Questionnaire (FAQ): 10-item questionnaire on instrumental activity of daily living, Everyday Cognition (ECog-12): scales to measure multiple cognitively relevant everyday abilities, covering six domains. BMI: Body Mass Index.

B. Patient Representation

We represented patient’s visits in a temporal mode and used them in our streaming clustering model. To incorporate temporality, we extracted the features in Table 1 to represent each visit for each patient. Age is grouped into four categories (Age ≤ 64, 65 ≤ Age ≤ 74, 75 ≤ Age ≤ 84, Age ≥ 85). Body mass index (BMI) is also divided into 4 groups (Underweight: BMI ≤ 18.5 kg/m2, normal: 18.5 ≤ BMI < 25 kg/m2, overweight: 25 ≤ BMI < 30 kg/m2, and obese: BMI ≥ 30 kg/m2). We also divided the years of education (Education ≤ 12, Education > 12) and the Epworth Sleepiness Scale (ESS) (ESS < 10, ESS ≥ 10) into two categories. For CU patients, we used all the visits of each patient. However, for dementia patients, we extracted the same features for all the visits prior to the dementia diagnosis visit. For example, suppose a patient has 7 visits in our cohort, and the dementia diagnosis was at visit 5. In that case, we use data from the first 4 visits in our model to predict dementia using these 4 visits. The same method was applied to MCI prediction, where we used all the visits prior to the MCI diagnosis. Each patient is labeled as dementia or CU. We also label the individual visits for each patient as MCI or CU. The resulting dataset and labels will be used in all the experiments in this paper.

C. Streaming Clustering Model

MUSC is a streaming clustering algorithm based on a Gaussian mixture model (GMM), coupled with possibilistic clustering which is an adaptive system for analyzing streaming multi-dimensional activity feature vectors with the goal of identifying signs of early diseases [25, 30]. The system is based on temporal analysis, including outlier detection, customization and adaption to new changes, together with the creation of new components for the (GMM) in the case of new emerging patterns. Since the number of groups in a streaming dataset is not usually known a priori, number of initial clusters in MUSC needs to be automatically determined. For this reason, possibilistic C-Means (PCM) [31] and Automatic Merging possibilistic Clustering Method (AMPCM) [32] are combined together to cluster the initial data points, detect anomalies and initialize the GMM.

After initializing the streaming clustering model using a small percent of the data, the remaining data samples are fed to the model in a streaming fashion (one sample at a time). When a new sample (x_{n + 1}) arrives, the Mahalanobis distance to the mean of each GMM is computed using equation (1). If the new data is closest to a particular cluster and the distance falls into a prespecified threshold, then this instance will be incorporated into that Gaussian. The threshold is the average distance between the data samples that were assigned to the existing clusters. When a new data sample is incorporated into the existing clusters, the threshold is updated incrementally. If the distance of (x_{n + 1}) is greater than the threshold, the newly arriving data sample is identified as an outlier and moved to the anomaly list. The parameters of each cluster (mean, covariance, cardinality) are updated incrementally each time a new sample is added to that cluster using equations (2) to (4). The points in the anomaly history may or may not indicate the emergence of a new cluster. The changes in the anomaly list are monitored in two different ways. First, we measure the Mahalanobis distance between the points in the list and updated prototypes to check if any point could fit in one of the existing clusters. Points are incorporated in their nearest Gaussians if they fall within a pre-defined range (the cluster has “expanded” to contain what was an outlier before). Second, we look for multiple emerging structures by clustering the outliers following the same procedure used to initialize the GMM at the beginning. For more details about the basic algorithm details and initialization steps, refer to [25].

$$d=\sqrt{{(x-\mu )}^T{\mathrm{\Sigma }}^{-1}(x-\mu )}$$ (1)

$$|{\mu }_{new}|=|{\mu }_{old}|+1$$ (2)

$${\mu }_{new}={\mu }_{old}+\frac{x_{n+1}-{\mu }_{old}}{|{\mu }_{new}|}$$ (3)

$${\mathrm{\Sigma }}_{new}=\frac{(|{\mu }_{new}|-1)\ast {\mathrm{\Sigma }}_{old}+{(x_{n+1}-{\mu }_{old})}^T(x_{n+1}-{\mu }_{old})}{|{\mu }_{new}|}$$ (4)

where μ_new, |μ_new|, and Σ_new are the new mean, cardinality, and covariance matrix of the updated cluster respectively.

Streaming clustering models are usually used to look for trends in the data, and recently they have been introduced as an online classification technique for streaming data. In this paper, we used MUSC to look for distinct groups in the MCSA cohort and to study the association in the clusters. In addition, we introduced heuristics-based approach on MUSC’s output to predict the possibility (i.e., early alert) of a patient to progress to MCI and dementia. MCI prediction is tracked through monitoring the cluster changes between consecutive visits of the same patient. If the current visit is assigned to a different cluster compared to the previous one, we consider that a sign of change in the cognitive status (i.e., CU to MCI). Therefore, every time the model assigns a health visit to a different cluster compared to the prior visit, we detect that change and we consider that visit potentially an MCI visit. On the other hand, dementia visit is detected through monitoring the anomaly list of MUSC. If any visit is flagged as an anomaly, we consider that abrupt change as a possible dementia visit. The pseudo code for predictive heuristics is displayed in Figure 1.

Figure 1.

Heuristics for MCI and dementia prediction.

D. Experiments

To evaluate the effectiveness of the proposed model, we first analyzed the clustering performance and the characteristics of each cluster. We also assessed the ability of the model to early detect the signs of MCI and dementia using sensitivity and specificity. In addition, we examined accuracy, precision, recall, and F1-score as a performance metrics. F1-score is an effective metric for imbalanced data with focus on the positive class by combining precision and recall. We compared the model performance to a long short-term memory (LSTM) model developed in [33] and evaluated on the MCSA cohort. Missing values were imputed using the mean of each feature.

III. Experimental Results

A. Cluster Analysis

MUSC was initialized with 4% of the total visits in the cohort. The visits of each patient are fed to the model in a streaming fashion (one visit at a time). Each visit, either assigned to one of the initial clusters or moved to the anomaly list. The clustering results are shown in Figure 2. Cluster 4 represents the visits that MUSC flagged as anomalies. Most of these visits are from dementia patients as in Figure 2–b (blue) or MCI visits (blue) as in Figure 2–a. Cluster 1 and cluster 3 also have more visits from dementia and MCI patients than those from CU patients. On the other hand, we notice that visits from CU patients are dominant in cluster 2 whereas there are more dementia and MCI visits in the other clusters (1, 3, and 4).

Figure 2.

Cluster distribution for a) MCI and CU visits, b) dementia and CU visits.

To better understand the characteristics of each cluster, we examined some of the features used in the model (Table 1) that shows different trends among the clusters. Figure 3–a displays the age distribution (Age ≤ 64, 65 ≤ Age ≤ 74, 75 ≤ Age ≤ 84, Age ≥ 85) in each cluster. It can be noticed that cluster 2 has more visits from a relatively younger population because the majority of the visits in this cluster are from healthier patients. However, the remaining clusters have more visits from older patients because these clusters have more visits from dementia or MCI patients. Figure 3–b displays the distribution of the ESS feature (ESS < 10, ESS ≥ 10) known to be associated with dementia [34]. Clusters 1, 3, and 4 have overall more visits from patients with ESS ≥ 10. On the other hand, cluster 2 has more visits from patients with ESS < 10, which meets our expectation that CI patients tend to sleep more during the day (larger ESS value). The sex distribution is depicted in Figure 3–c, where we see more visits from female patients in the healthier group (cluster 2) and more males in the CI clusters (clusters 1, 3, 4). Finally, in Figure 3–d, we investigated the education feature in the clusters. Participants with education ≤ 12 years are mainly dominant in clusters 1, 3, and 4, while visits from the population with education > 12 years are primarily in cluster 2, showing meaningful correlation between low education level and CI.

Figure 3.

Clusters Characteristics for: a) Age, b) ESS, c) Sex, d) Education

B. Examples of Patient Characterization in Clusters

To further investigate the model performance, we examined multiple dementia patients in the cohort. In each patient, we compared the features of the visits at which the model detected a possible dementia case and compared them to the features from the previous visit to better understand the main contributing features. The visit that was flagged as an anomaly by the model is considered a dementia visit due to the drastic change from being part of the clusters to being an outlier. It should be noted that the value of the features is after normalization, so it may not reflect the actual feature value in the cohort. The first patient case is in figure 4. We can see that features (23: Hallucinations, 28:Apathy, 30:Irritability/lability, and 36: personal care) are the main ones that led that visit to be flagged as an outlier and predicted as a dementia visit. Here, the dementia prediction visit is the visit before the actual dementia diagnosis in the MCSA cohort. The second patient case shows that features 7 (alcohol-related feature) and 37 (FAQ: 10-item questionnaire on instrumental activity of daily living) are the main drivers of flagging that visit as dementia compared to the previous visits. In the third patient, we see a notable increase in multiple features, 22, 25, 27, and 32, all of which are neuropsychiatric characteristics. The fourth patient case has features 29, 32 (neuropsychiatric features), and 36 (personal care feature). These cases show an aspect about the model behavior in flagging a visit as dementia where one or multiple features can contribute towards that. It also adds an explainability to the model. We can notice that neuropsychiatric features have a large impact on the model decision to flag a visit as an outlier (a possibility of dementia). This analysis shows an opportunity to track increasing or decreasing trends, which can be employed to generate early warnings for CI.

Figure 4.

Case studies comparing the features value before and at the visit of Dementia prediction

C. Prediction using Streaming Clustering

Streaming clustering has been proposed as a streaming classification model [35]. Even though the streaming clustering is unsupervised and not strictly intended for prediction, we explored its capability in predicting MCI and dementia.

To further evaluate the model performance in MCI and dementia prediction, we used sensitivity and specificity as evaluation metrics. MCI prediction is tracked through monitoring the cluster changes between consecutive visits of the same patient. If the current visit is assigned to a different cluster compared to the previous one, we consider that a sign of change in the cognitive status, and hence flag that visit as a potential of MCI. Dementia, on the other hand, is detected through monitoring the anomaly list of MUSC. If a visit is flagged as anomaly, we consider that sudden change as a possible dementia visit. Table 2 shows the performance in predicting dementia and MCI visits. It can be observed that the model is performing slightly better in detecting the MCI visits than dementia visits. Flagging a visit as an outlier requires that visit to be completely different from the existing prototypes. In contrast, visits could be assigned to other clusters based on the closest prototypes, which may lead to better performance on the MCI cases. Therefore, more parameter tuning in the future could enhance the model performance in detecting dementia compared to MCI.

Table 2.

The sensitivity and specificity of the model

Dementia
Sensitivity	0.56
Specificity	0.75
MCI
Sensitivity	0.60
Specificity	0.80

We also compare the prediction performance against the LSTM model that was developed to predict MCI on the MCSA cohort [33]. Even though the streaming clustering is based on unsupervised learning, it produced higher accuracy and F1 score (0.73 and 0.52) than the LSTM model (0.71 and 0.46). We believe that the enhanced performance of the streaming clustering is that it uses all individual visits to incorporate the temporal information to detect any change in the features space, leading to higher precision.

Since we predict dementia by monitoring the visits that are flagged as outliers, we examined how early our model can predict dementia prior to the actual diagnosis in the MCSA cohort. Figure 5 displays the visits at which the model correctly detected dementia prior to the actual diagnosis (the visits in the x-axis represent how early the model predicts dementia compared to actual dementia diagnosis; i.e., ‘dem v’ represent the actual dementia diagnosis visit in the MCSA, -v1 means 1 visit prior to actual dementia visit, -v2 means 2 visits prior to, and so on). More than half of dementia patients were detected (abnormal – i.e., outlier from the clusters) 2 or more visits prior to the actual dementia diagnosis. This shows the potential of applying our model as an early alert system to help clinicians detect early signs of dementia. It is worth mentioning that even though MCI is a leading precursor for dementia, it is not used in the model as input. Therefore, using MCI could enhance the model performance and leads to predict dementia in earlier visits.

Figure 5.

Distribution of the index at which the model flagged a visit as outlier (dementia prediction)

IV. Discussion

There is a lack of automated methods using temporal information in an unsupervised manner to predict CI in its early stages despite its widespread prevalence in the aging population. Early detection of CI could be a critical step towards better planning and treatment of AD and other forms of dementia. Therefore, we have explored an unsupervised learning model (temporal streaming clustering) to look for distinct groups in older adults. We compared the performance of the proposed model with other supervised learning-based techniques in the literature.

Our analysis demonstrated the ability of streaming clustering to capture potential MCI and dementia in early stages using temporal information of health visits. Our model can be applied for both MCI and dementia at the same time unlike most other models that either designed for MCI or dementia prediction alone. The model’s behavior can be explained as shown in patient examples (section III.B), which is an important aspect for explainable AI. Despite the use of unsupervised learning, our model outperformed a supervised LSTM-based model [33] with 6% increase in F1 score on the same MCSA cohort.

We noticed that different features led to flag a visit as dementia for different patients (Figure 4). It might reflect the heterogeneity of phenotypes and causation in dementia cases (e.g., AD, Lewy body dementia, vascular or mixed (e.g., AD/vascular), frontotemporal dementia etc.). Although we use cluster change as MCI, it may be a change in the cognitive status due to aging, not necessarily MCI. Therefore, further analysis of the clusters could improve MCI prediction.

The limitation of this study includes the use of variables in the MCSA that may not be readily available in other institutions. We plan to extend our model by extracting variables from routine EHRs to extend the applicability of the proposed method. The model was developed using the data in a single institution and thus the external validity will be needed to assess its generalizability.

V. Conclusions

A streaming clustering using a longitudinal visit data demonstrated a good potential in finding distinct older adult’s characteristics and predicting early signs of MCI and dementia. The model’s predictive performance in MCI and dementia was higher than the supervised learning (LSTM temporal approach). Unlike most existing MCI or dementia prediction, our model is based on unsupervised learning and tracks cognitive status changes from CU to MCI or to dementia at the same time, reflecting the nature of a slowly progressing disease over time. Our approach could facilitate early alert for MCI and dementia, addressing a significant current delay in diagnosis and thus improve treatment plans and health outcomes for patients.

Our next step is to apply the model with additional features from patient EHR documents. We also plan to initialize the model with publicly available data such as ADNI [23] and stream the data from the MCSA cohort to test the generalizability of the model. We will further investigate the features primarily contributed to the clusters to enhance model explainability.