Multi-level medical periodic patterns from human movement behaviors

Abstract

Human movement behaviors could reveal many interesting medical patterns. Due to the advances in location-aware devices, a large volume of human movement behaviors has been captured in the form of spatio-temporal trajectories. These spatio-temporal trajectories are useful resources for medical data mining, and they could be used to classify which trajectory passes through medical centres and which one does not. Traditional approaches utilise time-series datasets while ignoring spatio-temporal semantics in order to detect periodic patterns in medical domains. They also fail to consider the inherent hierarchical nature of patterns. We investigate a medical data mining framework that generates multi-level medical periodic patterns. A Geolife dataset is used to test the feasibility and applicability of our framework. Experiments demonstrate that the proposed framework successfully distinguishes those who periodically visit medical centres from those who do not, and also to find multi-level medical periodic patterns revealing interesting hierarchical medical behaviours. One potential application includes an automated personalised medical service. For instance, medical institutions can send personalised relative medicine information to people who regularly visit certain medical centres. It will be useful for the discovery and diagnosis of diseases for patients.

Introduction

With rapid developments in location-aware data acquisition gadgets, a massive number of spatio-temporal (ST) trajectories has been gathered and become available for data mining. A ST trajectory captures a trail of human movements, and the trajectory is spatially located and temporally recorded. Mining ST trajectories has become one of the hot research topics in data mining [14]. It delivers a new opportunity to analyse the behavior of human movements. It is a promising resource to distinguish those people who regularly visit medical centres for treatments (possibly patients) and for work (possibly health professionals) from those who do not. Therefore, it could be used to identify a set of patients or health professionals from massive trajectories in order to develop micro marketing or to further derive periodic patterns from these identified trajectories. For example, a person regularly visiting a medical place at 10am each Sunday for two months could be seen as a patient and a person periodically coming to the medical centre at 9am everyday could be a health professional working at the medical centre. Once these patients and health professionals are identified, then they can be further mined to reveal periodic patterns.

Mining ST periodic patterns [14, 17, 18] provides insights into ST trajectories, and it is to find periodic patterns from ST trajectories. Spatiality and temporality are too important aspects along with the sequential nature of ST trajectories. Due to the technical limitations and weather conditions, these trajectories are typically noisy, hierarchical and irregularly sampled. Thus, an effective preprocessing step to make them suitable for mining is a must in ST periodic pattern mining. ST periodic pattern mining could identify health-related people, and to detect multi-level medical patterns revealing valuable medical behaviours for those health-related people.

There are two main groups in periodic pattern mining in past studies: general periodic pattern mining, and ST periodic pattern mining. The former includes periodic pattern mining in event/sequence [4, 10], time series [2, 7, 9, 11, 24, 27, 28, 32] and social networks data [8, 21] whilst the latter involves ST trajectories [5, 14, 16–18, 29]. In medical contexts, most past studies in periodic pattern mining belong to the first type. They try to mine patterns from time series health datasets [2, 7, 11] and they are unable to mine patterns from ST trajectories. Some recent studies in medical domains attempt to utilise patient healthcare trajectories [22], to mine comorbidity patterns from outpatient records [3], and to analyse health conditions of elderly people using wearable devices [31]. However these studies are for utilising trajectories for specific pattern mining, and they are not for medical trajectory pattern mining. Recent studies [29, 30] are proposed to handle irregularly sampled and noisy GPS-collected trajectories, and to mine multi-level hierarchical periodic patterns. Here, we introduce a medical periodic pattern mining framework which is based on two recent ST periodic pattern mining methods. It is designed to classify a set of trajectories into positive trajectories exhibiting periodic visits to medical centres, and to find hierarchical medical periodic patterns. We propose two modified approaches based on our previous work [29, 30]: ST dominant approach and semantics-dominant approach. The former is based on trajectory clustering whilst considering the spatial and temporal aspects simultaneously in the process of PPM. The latter extracts semantic information from background information to obtain dense regions whilst taking spatial and temporal aspects into account at the same time. In this paper, the latter can extract medical centres from background maps for medical periodic pattern mining.

A list of main contributions are to:

• propose a multi-level pattern mining framework from ST trajectories which reveals medical periodic patterns;

• utilise two recent ST hierarchical periodic pattern mining methods to identify positive trajectories exhibiting regular visits to medical centres;

• identify single-level and multi-level medical periodic patterns;

• provide experimental results with a real world trajectory data, Geolife.¹

The rest of paper is structured as follows. Periodic pattern mining in medical contexts, and cutting-edge ST periodic pattern mining approaches are surveyed in “Periodic pattern mining” section. “Framework for medical periodic pattern mining” section proposes a medical periodic pattern mining framework and “Experimental results” section lists interesting medical periodic patterns, and presents comparative experimental results. “Conclusion” section concludes and presents directions for future work.

Periodic pattern mining

Definition

A term, ST trajectory, is used to depict a set of locations and corresponding time stamps for a moving object. It is represented by a triplet (lon, lat, t) where (lon, lat) represents a spatial location, lon represents longitude and lat stands for latitude, t represents a corresponding time stamp. A ST trajectory is a set $T=\{(x_1,y_1,t_1),(x_2,y_2,t_2),\dots ,(x_n,y_n,t_n)\},$ such that $t_i<t_{i+1}$ for all i $\in$ $\{1,\dots ,n\},$ each ($x_i,y_i,t_i$) is a trajectory node (a GPS sample point) at time $t_i,$ and $|t_{j+1}-t_j|\ne |t_{k+1}-t_k|$ for $\exists j,k$ where $1\le j\ne k\le n,$ where T is irregularly sampled because of bad weather conditions, device malfunctions, GPS reception errors, or unpredictable uncertainties. ST periodic pattern mining is to detect all periodic patterns from a given ST trajectory. Typically, it first identifies reference spots (where trajectory nodes are clustered), then finds periodic patterns for those reference spots.

Literature review

There are two major groups in periodic pattern mining: (1) general periodic pattern mining, and (2) ST periodic pattern mining. For general periodic pattern mining, Cao et al. [4] and Huang and Chang [10] detect periodic patterns from sequence data, and [2, 7, 9, 11, 28] mine time series data to detect patterns. In medical contexts, past studies in periodic pattern mining focus on time series datasets [2, 7, 11]. Ilayaraja and Meyyappan [11] introduces an approach to detect frequently occurring diseases in certain geographical places at a given time period using Apriori-based technique. Berlingerio et al. [2] uses time annotated sequences to detect associative frequent patterns and Froelich and Wakulicz-Deja [7] uses Fuzzy Cognitive Maps (FCMs) to find medical concepts from temporal diabetes data for periodic frequent pattern mining. Ismail and Hassan [12] proposes an algorithm named PPFP-growth (Productive Periodic-Frequent Pattern-growth) to discover all productive-associated periodic patterns from Body Sensor Networks (BSNs) to promote important decision making in healthcare. Ismail and Hassan [13] proposes an efficient PPFP mining algorithm to detect periodic patterns from a large data volume. One common problem is that they manipulate time series data with the temporal dimension, but do not consider the spatial dimension that represents ‘where’ periodic patterns take place. Also, these traditional approaches are limited to single-level patterns ignoring the inherent hierarchical nature of patterns.

The latter ST periodic pattern mining is a promising resource to find these ST associated periodic patterns. Several studies have been studied [5, 14, 16–18, 29, 30]. These past attempts can be grouped into three sub-categories: (1) the fixed period approach; (2) the reference spot approach; and (3) the semantics-based approach. The fixed period approach Cao et al. [5] makes use of pre-defined periods to detect periodic patterns. It combines bottom-up and top-down mining approaches to look for periodic patterns based on a user-specified fixed time period. It first partitions a long trajectory into sub trajectories using the user-provided period, and then uses DBSCAN [6] to group trajectory nodes to detect dense regions, and assigns class labels to these detected dense regions. This approach is not fully exploratory but confirmatory instead. It is unable to find periodic patterns with various periods. Also, it is not straightforward to extend this approach to irregularly sampled trajectories. It also fails to reveal hierarchical patterns, but instead is limited to single-level patterns.

The reference spot approach [17, 18] typically has three steps: (1) detection of reference spots; (2) corresponding period detection for each reference spot; and (3) mining periodic patterns for each reference spot. This approach utilises the Kernel-based method [26] to detect reference spots that are frequently visited by moving objects. It also uses a mix of Fourier transform and autocorrelation [1, 25] to find a period for each reference spot. Finally, it uses Kullback-Leibler Divergence [15] to associate corresponding periods to reference spots. The authors extended their previous study to deal with trajectories with missing data using linear interpolation and movement predictions [19]. One major issue with this approach is that it does not handle spatiality and temporality simultaneously. Another problem is that this approach takes regularly sampled trajectories as input, thus it is unable to handle irregularly sampled ST trajectories. Similar to the fixed period approach, this approach is limited to single-level patterns, and fails to detect multi-level periodic patterns. To overcome these drawbacks, Zhang et al. [29] proposes an algorithm, Traclus (ST) based on a ST trajectory clustering to find periodic patterns. Traclus (ST) takes both spatiality and temporality into account to find all ST concentrations, and their associated periodic patterns.

The semantics-based approach [30] tries to improve these problems from traditional ST periodic pattern mining. It uses semantic information extracted from background maps, finds spatially and temporally aggregated dense regions from irregularly sampled trajectories, uses Hidden Markov Model (HMM) to find semantically meaningful stops (places where an object or a user stays more than a user-specified threshold indicating the object or user is doing a meaningful behaviour such as medical treatment, surgery and consultation), uses Lomb-Scargle periodogram [20, 23] to detect periods for each semantically meaningful stop, and lastly detects periodic patterns for each stop. Differently from the ST dominant approach, the semantics-based approach deals with irregularly sampled trajectories and thus it is a promising candidate for our study.

Summary

Table 1 compares traditional periodic pattern mining approaches for medical patterns from ST trajectories. Traclus (ST) and semantics-based approach are able to detect hierarchical patterns and both equally and simultaneously consider spatiality and temporality, thus these two approaches will be used for this study.

Table 1.

A comparative table of traditional periodic pattern mining approaches

	Fixed period	Reference spot	Traclus (ST)	Semantics
Spatio-temporal	$✓$	×	$✓$	$✓$
Irregularity	×	×	×	$✓$
Hierarchical	×	×	$✓$	$✓$
Medical semantics	×	×	×	$✓$

Framework for medical periodic pattern mining

Figure 1 illustrates an overarching framework of proposed multi-level medical periodic pattern mining from ST trajectories. Basically, it consists of three main phases: input phase, mining phase and presentation phase. In input phase, the framework requires a set of GPS-collected ST trajectories, and in mining phase it utilises two cutting-edge ST periodic pattern mining approaches [29, 30] to positive trajectories, and their corresponding multi-level patterns. In the last phase, the framework presents medical periodic patterns for further analysis. A modified version of [29] for medical periodic pattern mining is called a ST dominant approach and a modified version of [30] is called a semantics-dominant approach. Please refer to Zhang et al. [29, 30] for more details.

Fig. 1.

Overarching structure of proposed medical periodic pattern mining from ST trajectories

Both algorithms below display modified versions for medical periodic pattern mining. Algorithm 1 depicts a modified ST dominant approach whilst Algorithm 2 shows a modified semantics-dominant approach for medical periodic pattern mining. Lines 8–11 in Algorithm 1 and lines 14–18 in Algorithm 2 detect reference spots that pass through or include medical centres for our study. In the ST dominant approach, line 2 shows that a trajectory interpolation process is inevitable to make trajectories with regular time intervals. Line 4 presents that different combinations of properties $⟨\text{Direction; Speed; Time}⟩$ obtain various reference spots for PPM. Line 6 displays that we extract medical centres from background maps. Lines 8–11 show that if a reference spot is close to a medical centre, we will match the reference spot to the medical centre and detect regular periods for this medical centre. In the semantics-dominant approach, line 2 shows that medical centres can be extracted from background maps. Line 3 displays that we apply HDBSCAN to find hierarchical reference spots. Lines 4–18 present that a few key steps for this approach, such as stop episode finding in lines 4–10, place matching in lines 11–13, and period detection in lines 15–19. Finally, periodic patterns can be obtained by these two approaches.

Experimental results

Dataset

In this paper, a widely used real Geolife GPS trajectory dataset is utilised for experiments. The dataset displays a sequence of timestamped locations of people, and it captures people’s outdoor movement behaviours such as visiting places, shopping behaviours, medical treatment behaviours, and general outdoor activities. Figure 2 shows two ST trajectories and locations of medical centres. One red trajectory shown in Figure 2a, recorded from September 26 in 2008 to October 10 in 2008, exhibits periodic visits to medical centres and it is referred to as a positive trajectory in this paper. The black trajectory, shown in Fig. 2b recorded from October 25 in 2008 to November 10 in 2008, does not exhibit periodic visits to medical centres and it is referred to as a negative trajectory. A set of medical centres in the study region is shown in Fig. 2c.

Fig. 2.

Visualisations of two user trajectories, and medical centres in the study region: a a positve trajectory; b a negative trajectory, c locations of medical centres

Efficiency

In this section, we compare the efficiency of the whole procedure for both methods based on the real dataset. Figure 3 displays the efficiency of the whole procedure for both methods. The ST dominant approach spends 1817.557429 s whilst the semantic dominant approach takes 5.060988 s for the whole procedure. Obviously, the latter is much more efficient than the former. The main reason is that the ST dominant approach needs interpolation to make irregular raw trajectories regular for subsequent period detection. Note that, the semantic dominant approach does not need interpolation since it is able to handle irregular trajectories for period detection.

Fig. 3.

Efficiency analysis between the ST dominant approach and the semantics dominant approach

Reference spots for positive and negative trajectories

In this paper, we are interested in positive trajectories exhibiting periodic visits to medical centres. These positive trajectories are potentially useful for medical decision making, thus reference spots for the red positive trajectory are analysed here as an example. Note that, Geolife trajectories are irregularly sampled, and a time interval of 10 s has been used to interpolate those irregularly sampled raw trajectories for the ST dominant approach to be comparable to the semantic dominant approach. As you can see from Table 2, the semantic dominant approach generates more reference spots than the ST dominant approach. These reference spots are used to generate periodic patterns, thus finding more reference spots will likely result in more pattern generation and less chance of missing true positives.

Table 2.

The number of reference spots for the positive trajectory shown in Fig. 2a

Approach	Number of reference spots
ST dominant approach	9
Semantic dominant approach	16

Medical periodic patterns for positive trajectories

We mine medical periodic patterns from positive trajectories using two algorithms under study, and present them to infer movement patterns and behaviors. Note that, background semantic information is not considered for reference spot detection in the ST dominant approach, thus we post-process detected reference spots to match with nearest medical centres with the ST dominant approach in order to compare its results with the semantic dominant approach. Figure 4 displays corresponding reference spots for the trajectories shown in Fig. 2. Using the ST dominant approach, 9 reference spots detected as in Fig. 4a for the positive trajectory shown in Fig. 2a, and 10 reference spots found Fig. 4b for the negative trajectory shown in Fig. 2b.

Fig. 4.

Detected reference spots with the ST dominant approach: a the positive trajectory in Fig. 2a; b the negative trajectory in Fig. 2a; c a zoomed area for (2,3,4,5,6) the green circle in (a); d a zoomed area for the blue circle in (b)

Figure 5 displays results with the semantic dominant approach for the positive trajectory depicted in Fig. 2a. In this paper, arrows and numbers are used to indicate ith reference spots. Figure 5a displays all reference spots, and Fig. 5b–d illustrate zoomed areas for the red circle, blue circle and green circle in Fig. 5a, respectively.

Fig. 5.

Detected reference spots with the semantic dominant approach; a the positive trajectory in Fig. 2a; b a zoomed area for the red circle; c a zoomed area for the blue circle; d a zoomed area for the green circle

Periodic patterns for the positive trajectory depicted in Fig. 2a with the ST dominant approach are shown in Table 3. Note that, the ST dominant approach aims at finding periodic paths, detected reference spots 6–9 are path-like patterns, and do not necessarily match with the medical centres depicted in Fig. 2c. As we can see from Fig. 4, the detected periodic pattern 6 → 0 → 7 and 9 → 0 → 8, reference spots 6–9 are parts of roads. These results demonstrate the unsuitability of the ST approach for medical periodic pattern mining.

Table 3.

Periodic patterns with the ST dominant approach for the positive trajectory

Reference spot	Period (h)	Periodic patterns
8	2	9 → 0 → 8
7	9	6 → 0 → 7

Three obtained periodic patterns with the semantic dominant approach is shown in Table 4. The first pattern, Peking University People’s Hospital  6 → 0 → Building 5, depicts a periodic pattern from Peking University People’s Hospital (reference spot 6) to Building 5 (reference spot 5). A period of 8 h is attached to reference spot 5 (a building). This explains that the user periodically goes to Building 5 (reference spot 5) every 8 h. Note that, 0 represents the moving object is not in any detected reference spot. The second pattern is a simple one indicating a periodic pattern from no known reference spot to Building 15 with a period of 3 h. The third pattern is: Student dormitory (reference spot 13) → 0 → Medical centre (reference spot 14). This medical periodic pattern explains the user goes to the medical centre from the student dormitory every day, precisely with a period of 23 h.

Table 4.

Periodic patterns with the semantic dominant approach for the positive trajectory

Reference spot	Period (h)	Periodic patterns
Building 5	8	Peking University Hospital 6 → 0 → Building 5
Building 15	3	0 → Building 15
Medical centre 14	23	Student dormitory 13 → 0 → Medical centre 14

These medical periodic patterns reveal this user’s health-related movements and regular behaviors. We can derive three possible inferences from these patterns. The first one: Peking University People’s Hospital 6 → 0 → Building 5, implies that the user might be a health-related researcher or student who works/studies in Peking University People’s Hospital and goes to Building 5 at a period of 8 h (might be a full time student). The third pattern implies that the user could be a student staying in a student dormitory: Student dormitory → 0 → Medical centre 14. The student goes to a medical centre regularly every day at a period of 23 h. It could be inferred that the student needs a light treatment everyday at the medical centre.

As we discussed above, the semantic dominant approach is better suited for medical periodic pattern mining under study. First, it classifies a user’s movement (trajectory) into a positive trajectory or a negative trajectory, and second it detects medical periodic patterns for the positive trajectory. These medical periodic patterns are solid suggestions for hypothesis generation or further inference analysis.

Hierarchical medical periodic patterns for positive trajectories

Figure 6 displays hierarchical reference spots obtained from the ST dominant approach. Figure 6a shows an obtained dendrogram which illustrates the hierarchical relationship between reference spots, such as reference spots 2, 3, 4, 5 and 6 can be merged as a reference spot (2, 3, 4, 5, 6).

Fig. 6.

Obtained dendrogram and hierarchical reference spots for the positive trajectory shown in Fig. 2a using the ST dominant approach: a dendrogram; b hierarchical reference spots

Table 5 shows identified hierarchical periodic patterns for the positive trajectory shown in Fig. 2a using the ST dominant approach. As mentioned earlier, this method fails to take background semantic information into account, thus detected reference spots are not necessarily matched with medical centres. In this case, two hierarchical periodic patterns detected are shown in Table 5. Since the ST dominant approach focuses on finding periodic paths, similarly with single-level reference spots, two hierarchical reference spots (2, 3, 4, 5, 6) and (8, 9) still do not match with the medical centres shown in Fig. 2c. For these two periodic patterns, two hierarchical reference spots are still parts of roads as shown in Fig. 6b. Thus, the hierarchical ST approach is not well suited for medical periodic pattern mining for our study.

Table 5.

Hierarchical periodic patterns for the positive trajectory using the ST dominant approach

Reference spot	Period (h)	Periodic patterns
(8, 9)	6	$7\to 0\to (8,9)$
(2, 3, 4, 5, 6)	4	$1\to 0\to (2,3,4,5,6)$

Figure 7a shows an obtained dendrogram for the semantic-dominant approach.

Fig. 7.

Obtained dendrogram and hierarchical reference spots for the positive trajectory shown in Fig. 2a using the semantic-dominant approach: a an obtained dendrogram; b–f hierarchical reference spots; g zoomed area of (5, 6) and (7, 8, 9)

Table 6 shows some meaningful multi-level periodic patterns that can match with certain medical centres. A hierarchical reference spot (15, 16) is comprised of a conference room (sometimes, it is used for informal conference), a supermarket and a restaurant. We can call this area non-working area. A single reference spot 14 is one of the medical centres in Peking University Health Science Centre. A hierarchical reference spot (10, 11, 12, 14) includes student dormitories, teaching building and medical centre 14, we can call this area the main activity area. A hierarchical reference spot (10, 11, 12, 13, 14, 15, 16) can be called the whole Peking University Health Science Centre. A single reference spot 6 represents the main building in Beijing People’s Hospital whilst a hierarchical reference spot (5, 6) is still a part of Beijing People’s Hospital. A hierarchical reference spot (5, 6, 7, 8, 9) can be called the whole Beijing People’s Hospital, including the inpatient department. A multi-level periodic pattern $14\to 0\to$ (15, 16) might show the user went to (15, 16) for food, shopping or informal meeting from spot 14 with a period of 6 h. In periodic pattern $13\to 0\to$ (10, 11, 12, 14), a single reference spot 13 is a laboratory, this is very normal that the user has a repeating behavior among the teaching building, laboratory and medical centre, which shows the user needs to have a repeating activity among the teaching building, laboratory and medical centre with a period of 12 h. We can call this area a medical area. A periodic pattern (5, 6) $\to 0\to$ (10, 11, 12, 13, 14, 15, 16) and (10, 11, 12, 13, 14, 15, 16) $\to 0\to$ (5, 6, 7, 8, 9) shows that the user has repeating activities between Beijing People’s Hospital and Peking University Health Science Centre with a period of 12 or 8 h. Note that, we do not show periodic patterns for hierarchical reference spots (7, 8) and (2, 3) because there are no periodic patterns for these two hierarchical reference spots. In addition, we can infer that the user is a student or teacher not a patient by periodic patterns with hierarchical reference spots. Interestingly, we cannot infer this with periodic patterns with single-level reference spots as discussed in the last section.

Table 6.

Hierarchical periodic patterns for the positive trajectory using the semantic dominant approach

Reference spot	Period (h)	Periodic patterns
(15, 16)	6	$14\to 0\to (15,16)$
(10, 11, 12, 14)	12	$13\to 0\to (10,11,12,14)$
(10, 11, 12, 13, 14, 15, 16)	12	$(5,6)\to 0\to (10,11,12,13,14,15,16)$
(5, 6, 7, 8, 9)	8	$(10,11,12,13,14,15,16)\to 0\to (5,6,7,8,9)$

Conclusion

Due to the advances in movement capturing devices, massive trajectories are being captured and available for data mining. A ST trajectory models a user’s movements and trails, and is a promising resource for medical periodic pattern mining. In this study, we investigate a periodic pattern mining framework for periodic medical pattern detection. We utilise two recent ST hierarchical periodic pattern mining approaches to find hierarchical medical periodic patterns, and demonstrate the usability and applicability of the proposed framework in medical settings using a popular real-world trajectory dataset.

Experimental results demonstrate that the proposed framework successfully classifies a user’s trajectory into either a positive trajectory (periodically visiting medical centres) or a negative trajectory (not periodically visiting medical centres), and also detects medical periodic patterns for the positive trajectory. The framework could be used to find potential customers (positive trajectories) from a large set of users (trajectories), and detected medical periodic patterns for positive trajectories could be used to reveal interesting patterns and used for hypothesis generation, targeted marketing, cause-effect analysis, and other data mining processes. In addition, we extend the single-level reference spots to multi-level reference spots for hierarchical periodic pattern mining. Experimental results demonstrate that some important and meaningful periodic patterns can be obtained in the presence of hierarchy. Future work includes extensive experiments with various datasets. They are required to further validate the robustness and feasibility of our framework. Preliminary experimental results show that the semantic dominant approach is a promising base for medical periodic pattern mining. However, it could be further enhanced by tightly-coupling semantic medical information into the framework.