Predictive and exposome analytics: A case study of asthma exacerbation management

1. Introduction

The future of healthcare is being shaped by the data revolution and the application of sophisticated data analytics towards improving preventive medicine and achieving personalized healthcare. While aspects of health data analytics such as the use of machine learning and statistical models have existed for decades [47], the new generation of health analytics incorporates a range of advanced and innovative technologies and tools such as (i) wearable monitoring devices and mobile applications for the capture of life-logging personal health data, (ii) bio-medical signal processing approaches for conditioning high-resolution data, (iii) “Big Data” paradigms to integrate the heterogeneous data, and (iv) data presentation capabilities to gain insights into large datasets [4,76]. This enables disease prediction and risk management in a ubiquitous personalized and continuous manner. The top motivational factor driving analytics in healthcare is reducing cost by leveraging predictive models to identify patients at risk for readmission so that healthcare providers can intervene early and make proactive decisions [11]. Prediction of patients adverse has a long history that has formed the basis for more recent innovations. For example, the study in [36] developed a chronic health evaluation system based on logistic multiple regression model to predict death rates of patients in sensitive care using data from medical records such as age, sex, indication of admission and severity of illness.

In 2015, the World Health Organization (WHO) produced a report stating that most of non-communicable disease deaths among under 70th age people was caused by four chronic diseases: i.e., cardio-vascular diseases, cancers, chronic respiratory diseases and diabetes mellitus [85]. Approximately 70%–90% of these chronic diseases and the elevation of incidence rates of these diseases are attributed to environmental factors. It is also estimated that the increasing patterns for many diseases such as asthma and degenerative diseases are marked more for environmental exposures than for genetic risks [30,80,81]. For asthma in particular, it is not sufficient to depend on bio-sensors to predict an attack because attacks result from mixtures of accumulated environmental exposure and personal triggers. It is necessary to analyze and discover the links between environmental exposures and asthma exacerbations. Exposure to environmental conditions varies across different locations and time intervals. This spatio-temporal nature of the exposure data best describes the dynamic behavior of individuals and enables the shift from static monitoring to spatio-temporally individual exposure assessment. Therefore, increased attention has been focused on the discovery and understanding of environmental influences on disease prevention [1,77,79], and the next wave of predictive health analytics will therefore involve aggregation of personal data with environmental data.

The assessment of personal exposure in space and time is affected by many variables with complex interactions among each other and the human system [9,49]. For example, small behavioral changes such as walking in different paths or staying farther way from traffic may cause significant changes on exposures to air pollute that exacerbates asthma. Thus, individual exposure assessment requires the recording of individual’s time-activity patterns besides the common environmental exposure scores. This necessity to embrace a holistic approach leads to the introduction of the exposome paradigm as a new research area to assess the impact of environmental exposures on human health [79].

The definition of exposome as the totality of environmental exposures indicates two conceptual dimensions with implications for measurement and analysis: (1) human’s health is affected by mixture of exposures; and (2) exposures occur across the lifespan [10]. This includes many factors to consider such as chemicals and pollutants, physical agents (e.g. noise), macro level factors (e.g. public health, schools, certain population density), and lifestyle factors (e.g. physical activity, sleep habits). Accordingly, exposome presents research issues in collecting, processing and analyzing exposure data to build a knowledge network for understanding exposures and the mechanisms underlying diseases. In this context, it is clear that using questionnaire to collect disease control data or to assess physical activities has become obsolete. In fact, several studies demonstrate that self-reported exposure data do not match the real-time monitored data [68].

Thus, developing effective predictive health monitoring systems requires integration of exposome analytical capabilities to facilitate knowledge discovery of environmental conditions individuals are exposed to in real time; and so detect the patterns of triggers and prevent environmental diseases. In this direction, a new line of interdisciplinary healthcare research and industry ventures has emerged to produce dynamic and real-time predictive health monitoring solutions [4,21,39,67]. While systems integrating exposome techniques with analytic capabilities hold substantial promises for tailoring care to individual patients and thus for transforming the future of healthcare, the penetration of such analytics solutions into actual clinical practice is in its infancy [11,61]. They are still presented by significant challenges in terms of data size, data scales, complex structures and relationships, uncertainty, and space and time constraints. Moreover, monitoring and assessment of individual exposure pose many challenges such as the limited resources and techniques to collect, process and analyze exposure data. In addition, currently there is no single chip to measure different types of exposures in a comprehensive manner for the “totality of environmental exposure” even for a short time interval. Moving this practice beyond its infancy will require, among other advances, novel statistical techniques capable of handling the totality of exposures and diseases in the context of interrelated disease states across short and long lifespan windows.

In this paper, we present an overview of a framework for combining exposome studies with predictive health analytics and its underlying concepts and essential processes. We also present a case study of a health analytics project that captures environmental exposures to monitor and manage risks of asthma exacerbations. Our work focuses on asthma, a common chronic respiratory disease affecting 235 millions of people of all ages in all parts of the world [85], since it is known to be highly affected by the environment. The development of asthma has been attributed to an inherent susceptibility towards a combination of many environmental variables [5]. While the list of asthma triggers include many variables (e.g. air pollution, weather, allergens, certain drugs or food, stress, and so on), the case study focuses on variables related to air quality and weather conditions. We propose computational models for the continuous learning of the relationships between environmental exposures and health outcomes, predicting health risks, and issuing meaningful and understandable personalized health feedback.

The remaining of this paper is organized as follows. In Section 2, we discuss exposome and predictive analytics on healthcare and highlight top open challenges. Section 3 discusses current methods of predictive analytics for asthma exacerbations management and their limitations. Section 4 presents our proposed framework of predictive and exposome analytics for predicting health risks of asthma, given patients and environmental exposure datasets. In Section 5, we present our investigation on a number of commonly used prediction models and discuss the limitation of the existing models. We then propose a two-step approach of regression models, which provides a meaningful and comprehensive feedback for patients. Finally, Section 6 presents future work, and provides some concluding remarks.

2. Analytics on healthcare

The term “Health Analytics” consolidates the various dimensions of “analysis” incorporating a range of technologies and tools: databases, electronic health record systems, data warehouses, web applications, clinical decision support systems and others that are integrated for interoperability and seamless processing of health data for insights [8,28,64]. The Healthcare Information and Management Systems Society, HIMSS (2013) defines health analytics as the “systematic use of data and related clinical and business insights developed through applied analytical disciplines such as statistical, contextual, quantitative, predictive, and cognitive spectrums to drive fact-based decision making for planning, management, measurement and learning” [29]. In a different and specific view, the term analytics in healthcare refers to a complete series of “integrated capabilities” that provide progressively deeper statistical insights into health-related information. These capabilities cover the areas of information access, quality, integration, storage, management, interpretation and governance [40]. Analytical capabilities in healthcare exist on a continuum of business-focused to clinical-focused solutions targeting three main objectives: improve organization performance, lower health care costs, and improve health or patient outcomes [14,40,64]. Generally, health analytics is perceived as a four-stage model composed of:

This discussion focuses primarily on the utilization of predictive analytical techniques in improving health outcomes and life quality. It also establishes the ground for “exposome analytics” as a joint capability to aid the understanding of exposure-disease relationships and predicting future risks as illustrated in Fig. 1 on page 531. Predictive analytics utilizes historical health data and medical knowledge to detect hidden patterns and predict health outcomes. On the other side, exposome analytics relies on latest advances in technology to track individual’s mobility patterns while simultaneously measuring the frequency and duration of exposure to environmental conditions in order to draw conclusions about related health impacts. Incorporating these conclusions or exposome knowledge discovery into health predictive modeling improves the performance of prediction and risk management of those diseases that are triggered by environmental exposures. In order to achieve such integration, we need to understand the conceptual background, processes, techniques and related issues of predictive and exposome analytics as we discuss next.

OPEN IN VIEWER

Fig. 1.

Integration of exposome and predictive analytics.

3. An overview of asthma exacerbation risk management

Asthma is a common chronic disease that affects some 235 millions of people of all ages in all parts of the world. While factors responsible for increasing asthma rates are not fully understood, the environment and lifestyle play the key roles to have caused the large increase in the numbers of people in the world with asthma. Asthma disease is characterized by recurrent attacks of wheezing and breathlessness, which vary in frequency and severity from person to person. Symptoms may occur several times in a day or week in affected individuals, and for some individuals become worse during physical activity or at night. During an asthma attack, the lining of the bronchial tubes swell, causing the airways to narrow and reducing the flow of air into and out of the lungs. Recurrent asthma symptoms frequently cause daytime fatigue, reduced activity levels, sleeplessness, and school and work absenteeism, all of which result in impairing quality of life. It is also a cause of substantial fiscal burden as it contributes significant costs to the health care system, particularly for (ED) visits and hospitalizations, which in many cases could be prevented. Actually, admission to hospital during an asthma attack indicates the first episode in the disease or a failure of preventive care for established asthma [85]. Thus, admission rates have been proposed as a target indicator for monitoring progress towards improved asthma care. This increases the demands for reducing asthma admission rates through a primary prevention system to reduce the level of exposure to common risk factors, particularly air pollution.

In the area of exposome analytics, research in [75] applied association rule mining to explore the effects of multiple air pollutants such as O 3 , PM 10 and CO on the risk of asthma in pediatric patients. Works in [52] developed an agreement coefficient statistic model (ACI) to classify personal GPS route data in order to generate time-activity categories for estimating personal exposures to air pollution while testing biomarkers of respiratory functions. A regression analysis model is devised in [18] to investigate the PM oxidative potential to induce airway inflammation in pediatric asthma patients using stationary air quality measurements. Similarly, works in [60,71] applied statistical modeling to estimate the effects of personal PM 2.5 exposure within different micro-environments (i.e. home, transit, school) in individuals with asthma utilizing real-time sensors and GPS data.

Unfortunately, while research has identified several environmental factors associated with the development of asthma, none has proven to be a causative agent. Further complicating the area of asthma management is the fact that most people diagnosed with asthma have isolated attacks separated by symptom-free periods. Thus, the goal of much asthma treatment is to control the condition at the individual level by reducing the risk of an attack. However, research has also shown that health responses to environmental exposure vary significantly over the population [50] and depend on many individual-specific variables, which makes managing health at the individual level challenging.

There have been strong-yet limited-research works to develop monitoring and risks management systems that capture individual-based measurement of exposure to certain levels of environmental factors for developing more accurate diagnoses on the causes of acute asthma episodes. In [43], a data mining alarm-system is proposed to predict asthma attacks automatically by mining asthma physiological symptoms (bio-signal data) and environmental factors together. The system collects environmental pollutant data, climatic and atmospheric data and all disease related data through the Internet and telecom communication. Then, data is analyzed by two methods: Pattern Based Decision Tree (PBDT) and Pattern Based Class-Association Rule (PBCAR). Both methods extract normal and high-risk sequential patterns as features; but then use different rule-mining techniques. The work in [46] developed a system for predicting asthma control deterioration one week ahead, by using self-monitoring tool, the Asthma Symptom Tracker (AST) assessments in conjunction with patient attributes and environmental variables. The AST assessment reflects the patient’s asthma control level over the past week. Each patient’s AST assessments were collected for six months; while environmental exposure data matched by time and location were obtained from multiple regional monitoring stations. Demographics and clinical status were also obtained. The system consists of two components: (1) a predictive model of the decision stump classifier, which makes a prediction based on a single independent variable; and (2) advanced predictive models built using multiple common classification algorithms and machine learning techniques. Similarly, the work in [19] used several existing data mining algorithms to predict asthma exacerbation utilizing tele-monitoring data such as physiological data and medication usage and identifies important attributes in asthma exacerbation.

A cloud-based predictive modeling of asthma readmission is also proposed in [12] using a set of classification and regression methods that utilize medical record information for a cohort of case patients with asthma readmission and matching control patients. The predictive module consists of several stages. First, the cohort construction and feature construction stages are conducted. The feature construction module allows the user to specify a method of aggregation for each event with respect to its values for each patient; while multiple techniques are used for feature selection (i.e. raw features, the Chi -square feature selection, F-value based feature selection). The predictive module then runs a combination of feature selection and classification tasks. The system was able to predict patients at risk for 12-month asthma readmission. The results of different predictive modeling tasks are also collected and displayed to the user in a highly intuitive, interactive user interface on a private server. Another predictive analytics approach to predict avoidable readmissions among patients is proposed in [19]. The proposed model incorporates patient’s history of readmission and impacts of patient attribute changes over time on a tree-based classification method to estimate the predicted probability of readmission. The data is collected from patient files then aggregated according to different risk factor categories (i.e. demographics, socioeconomic). Two types of predictive modeling are then applied: (1) classification and logistic regression models to focus on readmission occurrence; and (2) Cox proportional hazard model to detect time-to-readmission patterns.

In a different direction, several research works present pattern recognition models to identify consequential relations between environmental factors and asthma attacks. The work in [33] presented a pattern recognition model to identify consequential relations between environmental factors and asthma attacks. The model detects the prevalence of asthma by finding complex interrelations between air pollution, weather, and asthma exacerbation. Once a burst in asthma related message is detected, historical physical sensory data is analyzed, to find a set of complex risk factor patterns that might have resulted in the burst. A brute force algorithm is used to generate all possible combinations of events in a pattern and different time lags between them; while a finite-state automaton method is applied for each pattern to count the frequency of a set of candidate patterns. The results indicate that recognizing the time lags between different events (e.g. rain, temperature, PM 2.5 , wind) is an important aspect of correlation among multiple risk factors and asthma outbreak. In this same line, the work in [16] also uses a pattern recognition method to detect the prevalence of potential asthma risks in a specific spatio-temporal context, and to support an interpretation of a correlation between physical and social sensory data and asthma risks, in short exposure and real-time modes. The proposed method extracts related features then uses supervised learning approaches (e.g. SVM, association rule mining) to construct the event detector.

The work in [57] identifies to which extent heterogeneous information contributes towards the prediction of asthma, while comparing linear (i.e., logistic regression) and non-linear (i.e., random forests) machine learning predictive models fitted with different feature combinations. The findings indicate that random forests algorithm appears to be an effective tool to extract and integrate large numbers of heterogeneous predictors/attributes simultaneously for asthma attack predictive models. Regarding the importance of features, the findings confirm the important contribution of allergen sensitization (i.e. dust mite, dog, cat), along with lung function markers, in predicting asthma diagnoses or symptom patterns. A logistic regression based predictive model is also proposed in [48] using the Asthma Control Questionnaire (ACQ) score over time to predict risk of a future asthma exacerbation. ACQ includes nighttime waking, symptoms on waking, activity limitation, shortness of breath, wheeze, rescue short-acting medication use and the scores of lung function test FEV (forced expiratory volume). The association between baseline ACQ score and exacerbation was assessed using a Cox proportional hazards model [73] adjusted for treatment assignment. A second Cox proportional hazards model used all ACQ scores before exacerbation as time-dependent covariates. The time-dependent model was selected to enable the use of all the ACQ scores over time before the exacerbation occurred, rather than the single baseline or the last ACQ score. The model indicates that time-dependent values of attributes (i.e. AM-PEF vs PM-PEV) are differently associated with the risk of asthma exacerbation, highlighting that morning peak flow and diurnal variation of peak flow were significantly associated with the risk of asthma exacerbation. The study also demonstrates a hazard rate (HR) trend that suggests that higher FEV1 values are associated with decreased risk of exacerbation. Such findings are confirmed by another multivariate logistic regression predictive model proposed in [41]. The study also indicates that changes in lung function gave the highest sensitivity and specificity as a proxy for exacerbations; while a combination of 20% decrease in PEF or 20% increase of symptoms (on 2 consecutive days) defines a sever exacerbations that can serve as a signal for the patient to contact the healthcare provider.

A recent study to estimate environmental effects on asthma exacerbations is presented in [42]. The study performed a cross-over analysis using variables about the weather and air pollution at 1-week intervals between cases and control before and after ED visits. The study acquired the weather and air pollution variables over a 6-year period as well as the data of patients who visited ED with asthma exacerbations over the same period. The results indicate that the high wind speed and low humidity were particularly associated with an increased risk of asthma exacerbations. More mathematical and statistical models to predict future risk of asthma exacerbation are presented in [13,72]. Research in [6] presents promising results in building a continuous real-time profiling and monitoring of individuals. The proposed method builds the “Voronoi map” where each Voronoi cell represents a region that has similar environmental conditions and calculates each individual’s exposures. The work in [32] follows up to present a Bayesian framework for estimating path selection probabilities from extremely sparse GPS data for the purpose of estimating a “measurement of interest” that varies with path and travel time.

In many of the above studies, one of the primary health indicators of lung function that is used in the management of respiratory disorders such as asthma is the peak expository flow (PEF) measurement. This measurement measures the degree of obstruction in the airways and monitors a patient’s ability to breathe out air. To interpret the significance of PEF measurement, a standardized “normal” value is often used, which is obtained from a chart comparing the person with asthma to a general population without breathing problems. One such chart is the EU scale, which is based on an individual’s gender, age and height [15]. On this scale, a male with the age 30 and 180 cm height should show his normal PEF as 635 L/min.

OPEN IN VIEWER

Fig. 2.

Kernel ridge regression.

A common approach to predict a patient’s next PEF value is to use ordinary linear regression of PEF values against recent environmental exposures. This approach aims to predict the average PEF value that would be seen over many days with the same or similar environmental exposure values. It often provides unsatisfactory predictions in the area of asthma management [8,12] where doctors and patients are more often interested in studying PEF values in the high-risk zone (lower PEF value zone). Moreover, as illustrated by Fig. 2 on page 537, which shows the results of Kernel Ridge regression with polynomial degree = 1 for one patient in our study, the outputs of linear regression models also tend to underestimate the true variance of a patient’s PEF readings.

OPEN IN VIEWER

Fig. 3.

The framework of SCHAS incorporating predictive and exposome analytics to manage risks of asthma exacerbations.

Despite the progress mentioned above and many other asthma studies, the fact remains that while the majority of healthcare organizations are planning for or already implementing analytics, current efforts address mostly descriptive analytics rather than predictive analytics. Analytics budgets are spent mostly on IT infrastructure and systems designed to capture and process data, and few resources have been devoted to real-time data analytics applications. As mentioned above, well-known classification algorithms and machine learning techniques have been used in exposome and predictive analysis, but the outcomes from these systems have shortcomings in explaining the relationships among many variables. Hence challenges remain in practical use of this analysis.

In this section, we present our ongoing work on Smart and Connected Health Alert System (SCHAS) project to develop a health monitoring system for modeling individual exposure to environmental triggers of asthma and our approaches to integrate exposome and predictive analytics.

4.3. Predictive analytics

In our study, we analyzed 9 datasets of five asthma patients who have been consulted and monitored by doctors and medical practitioners at Soonchunhyang University Bucheon Hospital, South Korea. The PEF values of each individual were collected twice a day (in the morning and in the evening) and the durations of the data collection vary from 364 days to 689 days. Each individual’s medical profile and the AM and PM data are shown in Table 3 on page 540 together with the “normal” values from the EU scale [15].

4.3.1. Health risk zoning

In order to further classify risk of respiratory distress, the American Lung Association classifies PEF values into 3 zones of measurement [3]; green, yellow, and red zones. These zones are called “Normal Zones” because they are established by studying PEF values over the entire population and they are commonly used to develop management plans as shown in Table 4 on page 540 (a). However, as Fig. 4 on page 541 reveals for three individuals in our study, this zoning can be problematic because the PEF values of some affected individuals are mostly in the red zone while the majority of the PEF values of other individuals are mainly yellow.

OPEN IN VIEWER

Fig. 4.

PEF data distributions of three individuals based on the normal zoning [15].

OPEN IN VIEWER

Fig. 5.

PEF data distributions of three individuals based on the quantile zoning ( a = 20 % , b = 80 % ).

High variability within the population of individuals’ typical PEF ranges causes difficulty in using the Normal PEF zoning scale for predicting health risk, especially for at-risk individuals. Hence, we propose a new individual-based zoning of asthma health risks based on quantiles as described in Table 4 on page 540 (b), which we refer to as “Quantile Zoning”. Quantiles are cut point values dividing the entire data set of a patient’s PEF readings into zones based on the percent of readings above and below the cut point. In the proposed zoning, the patient and doctor select a percentage, a, and the readings representing the lowest a % of PEF values represent the red zone. Data between a % and b % are classified as the yellow zone, and data in the top b % are in the green zone. The PEF values that separate these zones are then taken as the zone cutoffs in lieu of the American Lung Associations cutoffs, which are based on healthy individuals. Although we used 3 zones in our analysis, the number of zones in the quantile zoning can be a user input to the system. Figure 5 on page 541 shows the PEF data distribution of three individuals with each data point labeled based on the quantile zoning.

While the quantile zoning method seems to work better for asthma patients than other scales, yet there are challenges. In particular (1) optimal cut-off values for the zones vary by individual and can be difficult to establish, and (2) the values close to the boundaries between two zones tend to be noisy and degrade model performance. Therefore, we recommend that the quantile zoning method be used as a guideline or starting point for doctors and patients in selecting individual-based cutoff values for the zones.

4.3.2. Health risk probability

One useful measure of health risk in medical patients is the probability that an individual’s health level drops below a threshold, such as the yellow-red cutoff established above. In the context of asthma management, we let C PEF be a critical PEF value that is determined by medical practitioners and a patient based on his or her medical history, such as hospital visits, symptoms, adverse episodes, and medication use. The patient’s health risk can therefore be identified as the probability that his or her PEF value falls below C PEF , P ( y < C PEF ) . Even though C PEF may have been established based on the lower, say, 20 % of the patient’s PEF values, the probability of falling below C PEF on any specific day depends on many changing factors so can therefore be expected to differ from 20 % , and this probability can be used as a measure of a patient’s daily health risk.

There can be significant difficulty in doctors and patients arriving at a practical and useful interpretation of the risk associated with a given probability. For example, if a system issues a warning message that my PEF measurement will be below my critical PEF value with 20 % probability, the formal meaning is P ( y < C PEF ) = 0.2 . Practically, does this mean that I will struggle only slightly with my condition because the probability that my PEF value will be above my C PEF is 80 % ? Another challenge in the use of the risk probability is how to deal with many false warnings, because if patients are warned when they have, say, 20 % chance of falling below C PEF then 80 % of the warnings they receive will be false alarms. However, most people would definitely like to know when they have a 20 % chance of having an ER visit on a particular day.

4.3.3. Central health management problems

Asthma patients are often most interested in predicting either what risk zone they will be in the next day or in predicting the risk that their health will deteriorate below a critical threshold level. The challenges can be formulated as two statistical problems that can guide work in predictive and exposome analytics of asthma risk management.

Problem 1.

Determine the health risk level (red, green or yellow) of an individual’s next PEF measurement based on his/her recent exposures.

Problem 2.

Estimate the probability that a patient will experience a critical health state given his or her recent exposures.

For the remainder of the paper, we survey several methods for addressing these problems and we propose a twofold process for solving the second that mitigates the prevalence of false warnings by providing readily understood gradations for the risk level.

5. Prediction modeling and evaluation

5.1. Classification modeling

We now survey the most popular data mining methods [7] to solve the classification problem of the prediction of health risk zones and, in the context of asthma management, present the analysis results of five methods: (1) Decision Tree Gini (DT-G), (2) Decision Tree Entropy (DT-E), (3) Support Vector Machine (SVM), (4) Gradient Boosting (GB), and (5) Logistic Regression (LR). These methods employ divers modeling approaches and come forward with a set of advantages and drawbacks. Additionally, we discuss methods for mitigating challenges presented by the fact that we often have much more data in the yellow zone than in the red or green.

5.1.1. Classifiers with imbalanced classes

One of the most common challenges faced when developing a classifier is the class imbalance problem. A dataset is considered imbalanced if the class of interest (low PEF values or high PEF values) is relatively rare as compared to the other classes [45]. As a result, the classifier can be highly biased toward the majority class. A number of sampling approaches, ranging from under-sampling to over-sampling, have been developed to solve the problem of class imbalance [31]. One challenge with sampling strategies is determining how much sampling to apply. An over-sampling level must be chosen so as to promote the minority class, while avoiding overfitting to the given data. Similarly, an under-sampling level must be chosen so as to retain as much information about the majority class as possible, while promoting a balanced class distribution [45].

In this paper, we explore the above classification methods to predicting a quantile PEF zone of the individual based on his/her 24 exposure variables. Let class 0 , class 1 , and class 2 correspond to the quantile Red zone, Yellow zone, and Green zone, respectively. We apply the combination of over- and under-sampling methods [45] to build our prediction models and present the quality of these models using several performance metrics. The experiments on the quality of classifiers to predict asthma risk zones were conducted on 9 datasets of the five individuals’ daily PEF values and their exposure variables as described in Table 2 on page 540 and Table 3 on page 540.

OPEN IN VIEWER

Table 5

Confusion matrix

	Predictive class i	Predictive ¬ class i
Actual class i	TP	FN
Actual ¬ class i	FP	TN

OPEN IN VIEWER

Table 6

Model performance: average values of metrics (3 quantile zones using a = 20 % , b = 80 % )

Dataset	Model	Accuracy	Precision	Recall	F 1	Dataset	Model	Accuracy	Precision	Recall	F 1
AM I 1	DT-G	0.59	0.53	0.59	0.53	PM I 2	DT-G	0.57	0.49	0.57	0.49
	DT-E	0.57	0.48	0.57	0.49		DT-E	0.56	0.46	0.56	0.47
	SVM	0.58	0.39	0.58	0.45		SVM	0.52	0.46	0.52	0.48
	GB	0.50	0.42	0.50	0.45		GB	0.49	0.41	0.49	0.44
	LR	0.56	0.47	0.56	0.48		LR	0.57	0.45	0.57	0.47
AM I 5	DT-G	0.56	0.48	0.57	0.48	PM I 4	DT-G	0.56	0.47	0.57	0.48
	DT-E	0.54	0.46	0.55	0.47		DT-E	0.55	0.42	0.56	0.45
	SVM	0.60	0.36	0.60	0.45		SVM	0.60	0.36	0.60	0.45
	GB	0.54	0.49	0.54	0.50		GB	0.49	0.41	0.49	0.44
	LR	0.58	0.48	0.59	0.47		LR	0.56	0.45	0.57	0.46

OPEN IN VIEWER

Fig. 6.

Prediction confusion matrices (3 quantile zones using a = 20 % , b = 80 % ).

5.1.2. Classification model evaluation

The confusion matrix is a commonly used method for determining the quality of a classifier. In a binary confusion matrix, the quality of a classifier is evaluated based on the ability to distinguish “positive” examples from “negative” ones. The confusion matrix is shown in Table 5 on page 542, where TP is the number of positive data points correctly classified as positive, FN is the number of positive data points incorrectly classified as negative, FP is the number of negative data points incorrectly classified as positive, and TN is the number of negative data points correctly classified data as negative. Several standard evaluation metrics are calculated using these numbers [31]: (1) Balanced accuracy = TP 2 ( TP + FN ) + TN 2 ( TN + FP ) , (2) Precision = TP TP + FP , (3) Recall = TP TP + FN , and (4)  F 1 -score = 2 ∗ precision ∗ recall precision + recall . For classification problems with several classes, as we have here (Red = class 0 , Yellow = class 1 and Green = class 2 ), one overall quality metric is arrived at by calculating these numbers for each class independently and averaging the results. In this context, for example, TP is the number of data points in class i correctly classified as belonging to class i and, FN is the number of data points of class i incorrectly classified as not belonging to ¬ class i .

Average performance metrics of the five classification models using the quintile zoning are presented in Table 6 on page 543. The analysis results of the models show reasonable average evaluation scores for use in health risk prediction as compared to many models in health applications. For example, the average accuracy of the models was between 48% and 60%. These metrics, however, do not always represent the quality of the prediction models. This can be seen in the confusion matrix of each model presented in Fig. 6 on page 543. Although the proposed sampling method improved the models to handle imbalanced data distribution, predicting the minority class may result in high false risk predictions.

OPEN IN VIEWER

Table 7

Binary model performance: accuracy (2 quantile zones using a = 20 % )

Models	Dataset
	AM I 1	PM I 1	AM I 2	PM I 2	AM I 3	PM I 3	AM I 4	PM I 4	AM I 5
DT-G	0.74	0.77	0.78	0.76	0.73	0.77	0.76	0.77	0.78
DT-E	0.79	0.78	0.79	0.78	0.77	0.78	0.76	0.79	0.78
SVM	0.80	0.80	0.80	0.80	0.80	0.79	0.80	0.80	0.80
GB	0.73	0.74	0.77	0.77	0.74	0.76	0.76	0.73	0.75
LR	0.78	0.78	0.80	0.79	0.79	0.79	0.79	0.79	0.81

On the other hand, for more practical use of the classification in health risk prediction, the target zone can be the lower PEF value zone (higher risk zone) and this reduces the number of the classes to two, the area below a % and the area above a % . Our analysis results show that the performance of the models was improved and the average accuracy of the models was between 73 % and 81 % as shown in Table 7 on page 544. This brought our study to a two-step regression modeling approach that focuses on predicting the probability that the unknown PEF value is below the individual’s critical value C PEF .

Although the above performance metrics involve visualizations in terms of the confusion matrices that aid practitioners in integrating direct observations with theoretical models, the precise practical implications of the metrics can be difficult for non-specialists to understand. Indeed, metrics such as the F 1 score are sometimes more heavily relied on in the model development and training stage than on the actual model usage.

5.2. Regression modeling

We present a two-step regression modeling process based on logistic regression and quantile regression that aims to deliver easy to undertand and use predictions.

5.2.1. Logistic regression model

The logistic regression (LR) model is often used to estimate the probability of a binary response such as pass/fail or healthy/sick, based on one or more independent variables. It can also be used to analyze the effect that the presence of a risk factor has on the odds of a given outcome. Logistic regression is commonly used in medical fields to predict the risk of developing a given disease (e.g. diabetes or coronary heart disease), based on observed characteristics of the patient [20]. Logistic regression works by selecting the combination of coefficients for the linear combination z = a 0 + a 1 x 1 + ⋯ + a n x n of the n independent variables that best estimates the probability, p, of developing the disease when substituted into the logistic function p = g ( z ) = 1 / ( 1 + e − z ) .

From a clinical standpoint, logistic regression has the advantage of having outputs metrics that are relatively easily understood. For example a typical output would be “Based on your previous exposure, you have a 60 % chance of falling below your critical PEF value today.” One disadvantage is that a fairly low probability of a highly critical event definitely warrens attention by the patient, but this probability means that most such situations will be “false positives” in the sense that the critical event does not occur. Therefore, a second method of gaining more detailed information when the logistic model indicates a non-negligible probability of a critical event is important. Additionally, metrics evaluating the quality of a logistic regression model can be challenging to interpret, so a more easily evaluated method to be used when logistic regression indicates a high risk is desirable.

5.2.2. Quantile regression model

The Quantile regression (QR) model aims at estimating the conditional median or other quantiles of the response variable. In ordinary linear regression, one attempts to find the best linear equation y = f ( x 1 , x 2 , … , x n ) = a 0 + a 1 x 1 + ⋯ + a n x n that predicts the average value of the response variable y over all (theoretical) observations where the explanatory variables take on values ( x 1 , x 2 , … , x n ) . This is referred to as the “conditional mean” of the distribution of y given explanatory variables ( x 1 , x 2 , … , x n ) . Quantile regression on the other hand does not restrict attention to the conditional mean but rather attempts to model the quantile values of the conditional distribution, so it permits one to approximate the whole conditional distribution of a response variable. Since it allows investigation of values in the tails of the conditional distribution, quantile regression has recently found use in medical studies in which the more extreme values of a patient’s readings are of particular interest [26]. Specific details of the estimation process for quantile regression can be found in [37].

5.3. A two-step model

We propose a two-step prediction model combining logistic regression with quantile regression; Step 1: use logistic regression to predict a patient’s probability of falling below his/her critical PEF value and Step 2: If the probability given by step 1 is above threshold p determined by the patient and doctor, then use quantile regression to predict the PEF value associated with several quantiles, say 5, 25, 50 and 75, in the patient’s conditional PEF distribution to give the patient’s detailed information about the PEF values he/she is likely to encounter. Below, we evaluate the quality of each step of this model using a moving-window approach and show the results using the patients’ datasets.

OPEN IN VIEWER

Fig. 7.

Moving-window.

5.3.1. A moving-window approach

While numerous statistical tests exist for assessing the goodness of fit of logistic and quantile regression models [26,38], they are generally quite technical and their results are difficult to convey to patients for use in managing their health. Thus, we propose two new evaluation metrics, which yield immediately understandable information on the model’s performance to patients and practitioners. Evaluation of the models in both steps of our process was conducted with a moving window approach. This approach addresses both the number W of previous environmental exposures and PEF readings used to train the model and the number m of the future PEF predicted values for which the model is used before retraining. With advanced computing power and memory, updating the model every day is possible, but in real world applications saving computing power and memory are recommended unless more frequent updates on the model are required to maintain acceptable errors in prediction. Figure 7 on page 545 illustrates the use of the moving-window (W and m) in the analysis process. One might wonder why not take a larger training dataset, or even take all of the available data for training. The reason for avoiding this strategy is that it may result in loosing temporal information related to the fact that the relationship between health risks and environmental variables is dynamic and typically changes over time.

For any particular application of a model, the optimal values of W and m depend on many problem-specific factors including the desired model accuracy, computing power available and specific nature of the data. Therefore, the optimal values must be established during a period of training and testing in which the set of feasible pairs ( W , m ) is searched for the optimal values. In our experiments, we analyzed W = 30 , 60 , 90 , 120 , 150 , and m = 30 , 20 , 10 , 5 , 1 to find a good pairing.

5.3.2. Logistic regression model evaluation

In this section, we discuss only the pairing ( W , m ) = ( 90 , 30 ) we found to be optimal. The model was evaluated using the discrete binning technique. Since logistic regression outputs the estimated probability of a patient falling under a given threshold PEF based on the environmental exposures, the approach we take is to bin predictions together into intervals and then estimate the actual probability of falling below the threshold PEF level by comparing the number of days in each bin actually below the PEF to the total number of days in the bin. Specifically, we use bins [ 0 , 0.2 ] , ( 0.2 , 0.4 ] and ( 0.4 , 0.6 ] . The steps we follow are: (1) for each bin, identify the number of the PEF values in the test period for which the model predicts a probability within that bin, (2) for each bin, count the number of PEF values in that bin in which the observed PEF value is actually below the threshold PEF value C PEF , and (3) calculate the ratio of the numbers in (1) and (2). If the model is predicting accurate probabilities, the ratio in (3) is expected to be close to the midpoint of the bin’s interval.

OPEN IN VIEWER

Table 8

Logistic regression analysis (using critical PEF value based on the 20th percentile)

Probability	Synthetic data			AM I 1			AM I 3			AM I 5			PM I 2			PM I 4
	n c	N P	n c N P	n c	N P	n c N P	n c	N P	n c N P	n c	N P	n c N P	n c	N P	n c N P	n c	N P	n c N P
P [ 0.0 – 0.2 ]	46	3775	0.01	33	160	0.20	52	345	0.15	74	383	0.19	62	371	0.17	27	134	0.20
P [ 0.2 – 0.4 ]	32	107	0.30	9	30	0.30	13	38	0.34	12	49	0.27	19	40	0.41	8	51	0.16
P [ 0.4 – 0.6 ]	37	67	0.55	3	8	0.38	9	20	0.45	12	28	0.43	12	22	0.55	8	35	0.23
P [ 0.6 – 0.8 ]	54	81	0.67	4	12	0.33	6	17	0.35	10	28	0.36	7	24	0.29	7	19	0.37
P [ 0.8 – 1.0 ]	804	849	0.95	4	30	0.13	18	50	0.36	18	92	0.20	17	53	0.32	5	31	0.16

n c  = # of observations below C PEF , N P  = # of observations for which logistic regression predicts probability between a and b.

Assessing the quality of a model application by using a binning strategy can be difficult due to coarseness in the dataset not matching the bin cutoffs and also due to small bin sizes. Therefore, in addition to assessing the quality of the logistic approach on our real data, we also generated a synthetic data set of 5 , 000 observations. Each observation involved two explanatory variables x 1 and x 2 , which were selected from a uniform distribution between 1 and 10. For each pair ( x 1 , x 2 ) , we generated a random PEF value y from the normal distribution with the mean z = 5 + 8 ∗ x 1 + 7 ∗ x 2 and the standard deviation σ = 5 . Table 8 on page 546 shows the results for the synthetic data together with the results for five patients. The results of the synthetic data mostly follow the expected probability for each bin. On the other hand, the results of the patients’ datasets indicate that for probabilities under 60 % the predicted probability of falling below C PEF is reasonably close to the actual probability. For probabilities above 60 % , the model turned out to be less accurate, generally overestimating the probability of falling below the threshold.

5.3.3. Quantile regression model evaluation

To evaluate the quality of the QR model, we define a uniform measure of the relative error for each quantile τ and evaluated the model through extensive experiments on individuals’ datasets in a similar way to the method described for the logistic regression. For quantile regression, however, it is possible to define a bin-free measure of the relative error as follows. First we observe that for every combination of exposure values and for each quantile τ, the quantile regression model outputs the PEF value that is expected to be the location of the τ quantile for all PEF readings taken on days with the same or similar environmental exposure values. If the model is accurate, then we would hope that for any τ, approximately τ fraction of the test days actually has PEF value below that day’s predicted τ quantile PEF cutoff. Thus, a uniform measure of the errors can be calculated as follows: (1) Err τ = | N τ N − τ | , where N τ is the number of days having observed PEF value under that day’s predicted τ PEF quantile value and N is the total number of test days. For example, suppose that we had 200 test days, and the numbers of observed days below the daily PEF predictions for τ = 0.10 and τ = 0.20 were 21 days and 41 days, respectively, then the errors for τ = 0.10 and τ = 0.20 would be the same, 0.5 % , or Err 0.10 = Err 0.20 = 0.005 .

Our analysis shows that, as expected, smaller sizes training windows (i.e., W = 30 and W = 45 ) result in higher errors, but the results also show that increasing the training window size beyond a certain point does not always reduce the errors. Err τ of the model using 60 day or 90 day window are similar to or little higher than that of the model using 120 day or 150 day window. This means that a model can be developed for a patient in a relatively shorter period (i.e., in 2 months or 3 months) and the model can be refined further while the system serves the patient risk management. The optimal window size is varied depending on individuals. We present the relative errors for τ ranging between τ = 0.05 and τ = 0.95 using three patients’ datasets with varying window sizes in Fig. 8 on page 547. Similar results of the average Err τ between 3% and 6% were found in the analysis of other datasets.

OPEN IN VIEWER

Fig. 8.

Comparisons of relative errors of QR for varied window sizes.

OPEN IN VIEWER

Fig. 9.

Comparisons of average relative errors of QR for varied moving-window sizes.

With an optimal window of the QR model for each individual, we show the average relative errors of each individual’s quantile regression analysis for different window sizes while varying the size of the moving window m, as demonstrated in Fig. 9 on page 548. The results show that more frequent updates on the model reduces errors in the model but still the errors are reasonable when m = 30 .

5.3.4. Concluding evaluation remarks

To give a more intuitive understanding of the errors associated with our proposed method’s usage, we briefly imagine that patient I 2 were a patient using this model in practice. The PEF threshold for the logistic regression step was selected based on his/her overall 20th percentile PEF value, C PEF = 238 . The results in Table 8 on page 546 indicate that for probabilities under 60 % the predicted probability of falling below the threshold PEF is reasonably close to the actual probability. For probabilities above 60 % , the model turned out to be less accurate, generally overestimating the probability of falling below the threshold.

Once the patient knows his/her predicted probability of falling below the critical threshold, for small probabilities, the patient knows that in many such days he/she will be above the threshold, but sometimes he/she will fall below. In order to mitigate this uncertainty, the patient can turn to the results of the quantile regression and examine, say, the predicted 25th and 75th quantile cutoffs. Given the accuracy of the quantile model shown in Fig. 8 on page 547, these values would give the patient a fairly accurate picture of where to expect the middle 50 % of his/her days to lie in such a situation.

6. Conclusions and future work

The study of environmental exposures requires synchronization of environmental data, individuals’ moving trajectories, and the behaviors of individuals, such as physical activities, route selections, etc. By incorporating these data sources to characterize and predict path selection behavior, one can measure individual’s exposure to certain environmental conditions. The assessment of exposure enables analyzing the relations between negative health effects and levels of the environmental factors. In this paper, we introduced the concept of exposome analytics as a paradigm to capture the multitude of environmental exposures that impact human health and chronic diseases by using data mining techniques that identify relationships, associations and causality between environmental conditions and disease patterns. Integrating exposome analytics and predictive analytics on medical domain knowledge has the potential to predict and manage future risks; and to quantify the effects of prevention and intervention.

We presented a study of predictive and exposome analytics in the application of managing risks of asthma exacerbations. We discussed the conceptual framework of health risk modeling and environmental exposure assessment as a health monitoring platform for modeling individual exposure to environmental triggers of asthma exacerbation. Moreover, we utilized well-known data mining methods to develop a prediction model of the PEF risk zone and discussed the insight of the analysis. Restructuring the zones of the PEF values based on the individual’s data values can improve the practical usefulness of the modeling of classification of the PEF zones. Finally, we proposed a dual-regression model to predict the probability of a health risk. The proposed approach uses a novel exposome assessment paradigm that utilizes the spatio-temporal properties of the data in the model training process and hence results in improving the accuracy of prediction.

The research on health risk management that captures individual-based measurement of exposure is still open for more future work to solve topical challenges as measuring environmental conditions with a fine scale, treating uncertainties due to errors in device measurement and data sampling, integrating different representations of spatio-temporal data, and development of holistic visual representations for both structured and unstructured data from sensors, data structures and masses of streaming real-time data. With better data collection and exposome estimations, which can lead to more accurate regression models, the proposed framework has the potential to transform healthcare with a focus on self-management by issuing and monitoring individualized recommendations concerning patients daily routines. Therefore, one important future direction for work in exposome and predictive analysis should be to focus on spatio-temporal dynamic modeling of exposome. Accurate modeling of an individual’s behaviors and those behaviors’ effects on exposure in the presence of spatiotemporal uncertainty will be a key challenge to be overcome in the effort to measure individual exposome values accurately enough to allow significant improvements in the accuracy of individual health monitoring systems.

Open challenges also include the development of flexible and scalable data mining models to find accurate and sufficient information of moving objects’ patterns, behaviors and trends; and associated issues such as the parallelization of the computing model and availability of computing resources. For quantile regression modeling, we plan to develop a method to measure the quality of the model. We also seek to monitor more of asthma triggers and study their interactions as well as therapy adherence to understand their total impact on the improvement of asthma. We aspire that our work contributes to the diagnosis and prevention of asthma attacks; and to improve the quality of experience of patients and healthcare professionals.