Discovering Expected Activities in Medical Context Scientiﬁc Databases

Daniela D’Auria and Fabio Persia

Dip. di Ingegneria Elettrica e Tecnologie dell’Informazione, University of Naples ”Federico II”, Naples, Italy

Keywords:

Activity Detection, Scientiﬁc Databases, Cricothyrotomy Simulator, Medical Simulator.

Abstract:

Reasoning with temporal data has attracted the attention of many researchers from different backgrounds

including artiﬁcial intelligence, database management, computational linguistics and biomedical informatics.

More speciﬁcally, activity detection is a very important problem in a wide variety of application domains such

as video surveillance, cyber security, fault detection, but also clinical research. Thus, in this paper we present a

prototype architecture designed and developed for activity detection in the medical context. In more detail, we

ﬁrst acquire data in real time from a cricothyrotomy simulator, when used by medical doctors, then we store

the acquired data into a scientiﬁc database and ﬁnally we use an Activity Detection Engine for ﬁnding expected

activities, corresponding to speciﬁc performances obtained by the medical doctors when using the simulator.

Some preliminary experiments using real data show the approach efﬁciency and effectiveness. Eventually, we

also received positive feedbacks by the medical personnel who used our prototype.

1 INTRODUCTION

Reasoning techniques are very essential in many ap-

plication domains, such as video surveillance, cyber

security, fault detection, fraud detection and in clin-

ical domain, as well. In all cases, temporal infor-

mation is crucial. For instance, for what the clinical

research concerns, investigating disease progression

is practical only by deﬁnition of a time line; other-

wise, possible causes of a clinical condition have to

be found by referring to a patient’s past clinical his-

tory. In (Zhou and Hripcsak, 2007), the basic con-

cepts of temporal representation in the medical do-

main have been described in order to include: cat-

egory of time (natural, conventional, logical), struc-

ture of time (line, branch, circular, parallel), instant

of time vs. interval, and, absolute time vs. relative

time. Anyway, this is still a challenging and active

subject of research. The main goal of (Safari and

Patrick, 2013) consists in creating a special purpose

query language for clinical data analytics (CliniDAL)

to place in any clinical information system (CIS) and

answer any answerable question from the CIS. In

more detail, a category scheme of ﬁve classes of in-

creasing complexity, including point-of-care retrieval

queries, descriptive statistics, statistical hypothesis

testing, complex hypotheses of scientiﬁc studies and

semantic record retrieval have been designed to cap-

ture the scope encompassed by CliniDAL’s objectives

(Patrick and Cheng.Y., 2013). However, a review of

temporal query languages reﬂects that the importance

of time has led to the development of custom temporal

management solutions, which are mostly built to ex-

tend relational database systems (for instance, T4SQL

(Combi et al., 2007)). Many efforts in the relational

database ﬁeld have been conducted for developing ex-

pressive temporal query languages; nevertheless, they

still suffer from two issues: ﬁrstly, they are only ap-

plicable to structural relational databases; secondly, it

is difﬁcult for hospital staff with poor IT skills to ap-

ply them. On the other hand, in most ontology based

approaches composing queries can be difﬁcult due to

a complex underlying model representation and lack

of expressivity.

In other contexts, such as video surveillance, cy-

ber security and fault detection, the reasoning tech-

niques using temporal information are broadly used

for activity detection. Thus, several researchers have

studied how to search for speciﬁcally deﬁned patterns

of normal/abnormal activities (Hongeng and Neva-

tia, 2001). Vaswani et al. (Vaswani et al., 2005)

study how HMMs can be used to recognize complex

activities, while Brand et al. (Brand et al., 1997)

and Oliver et al. (Oliver et al., 2002) use coupled

HMMs. Hamid et al. (Hamid et al., 2003) use Dy-

namic Bayesian networks (DBNs) to capture causal

relationships between observations and hidden states.

Albanese et al. (Albanese et al., 2007) developed a

stochastic automaton-based language to detect activ-

ities in video, while Cuntoor et al. (Cuntoor et al.,

446

D’Auria D. and Persia F..

Discovering Expected Activities in Medical Context Scientiﬁc Databases.

DOI: 10.5220/0005146504460453

In Proceedings of 3rd International Conference on Data Management Technologies and Applications (KomIS-2014), pages 446-453

ISBN: 978-989-758-035-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2008) presented an HMM-based algorithm. In con-

trast, (Albanese et al., 2014; Albanese et al., 2011)

start with a set A of activity models (corresponding

to innocuous/dangerous activities) and ﬁnd observa-

tion sequences that are not sufﬁciently explained by

the models in A. Such unexplained sequences reﬂect

activity occurrences that differ from the application’s

expectations.

Other relevant works exploiting an events se-

quence deﬁnition are (Boselli et al., 2014a) and

(Boselli et al., 2014b): in particular, (Boselli et al.,

2014a) automatically identify cleansing activities,

namely a sequence of actions able to cleanse a dirty

dataset, which today are often developed manually by

domain-experts, while (Boselli et al., 2014b) describe

how a model based cleansing framework is extended

to address integration activities as well.

In this paper, we present a prototype architecture

designed and developed for activity detection in the

medical context. The context of use is very con-

crete and important, as it is represented by a cricothy-

rotomy simulator built by the BioRobotics Labora-

tory of the University of Washington, Seattle (USA)

(White et al., 2014; White et al., 2013; White et al.,

2012a). Such a simulator is very relevant in the surgi-

cal robotic ﬁeld, as it is very useful for helping med-

ical doctors when they are performing a cricothyro-

tomy. Our main aim consists in making the medical

doctors able to have a very fast feedback about their

performances when using the simulator, that is very

essential for this kind of applications. In order to do

that, we ﬁrst acquire data in real time from the simu-

lator, when used by medical doctors, then we store the

acquired data into a scientiﬁc database and ﬁnally we

use an Activity Detection Engine for ﬁnding expected

activities, corresponding to speciﬁc performances ob-

tained by the medical doctors when using the sim-

ulator. Thus, we model the expected activities with

stochastic automata (Albanese et al., 2014; Albanese

et al., 2011) and exploit the activity detection algo-

rithms presented by (Albanese et al., 2013), as tem-

poral information is essential.

The paper is organized as in the following. Sec-

tion 2 brieﬂy describes the context of use of our pro-

totype architecture, which is the cricothyrotomy simu-

lator designed by the University of Washington, Seat-

tle. Section 3 shows the model used for activity detec-

tion, while Section 4 describes the architecture of the

developed prototype. Section 5 presents some prelim-

inary experiments using real data. Eventually, Section

6 discusses some conclusions and future work.

2 CONTEXT OF USE: A

CRICOTHYROTOMY

SIMULATOR

Modern airway protocols involve many techniques

to restore ventilation including bag-mask-ventilation,

placement of a laryngeal mask airway, and intubation

with or without videolaryngoscope. In cases where

conservative measures fail or when contraindicated,

the only methods remaining to re-establish ventilation

may be surgical. In the developing world where de-

vices such as the videolaryngoscope may not be avail-

able, accurate knowledge and training in the creation

of a surgical airway may have a signiﬁcant impact on

patient outcomes.

A cricothyrotomy is a life-saving procedure per-

formed when an airway cannot be established through

less invasive techniques. Although performing such

a procedure seems relatively straightforward, studies

have shown that those performed in the pre-hospital

setting were mostly unsuccessful (Wang et al., 2011).

A review of 54 emergency cricothyrotomies found

that the majority of the procedures performed in the

ﬁeld were unsuccessful or resulted in complications

(King et al., 2012). A military team identiﬁed gap ar-

eas in the training of cricothyrotomy in emergency sit-

uations; these included lack of anatomical knowledge

including hands on palpation exercises, poor anatomy

in medical mannequins, and non-standard techniques

(Bennett et al., 2011).

Most of the unsuccessful attempts were due to

inaccurate placement, and incorrectly identifying

anatomy. If the anatomy is not properly identiﬁed,

it is unlikely that the procedure will be successful.

Further, a large review of emergency airway cases

found that emergency cricothyrotomies performed by

anesthesiologists were successful in only 36% of in-

stances (Cook et al., 2011). Although many reports

suggest that the success rate of surgical airway place-

ment is low, publications from advanced centers with

extensive training for airway protocols including sim-

ulation show that pre-hospital cricothyrotomy success

rates can be as high as 91% (Warner et al., 2009).

Studies such as this suggest that with adequate train-

ing, the success rate of cricothyrotomy can be dramat-

ically improved. Thus, an improved method of train-

ing needs to be provided for this rare, but life-saving

procedure.

For such reasons, the BioRobotics Laboratory

of the University of Washington, Seattle (USA) de-

veloped a low-cost cricothyrotomy simulator (White

et al., 2014; White et al., 2013; White et al., 2012a)

from readily available components that is equipped

with inexpensive sensors. The simulator emphasizes

DiscoveringExpectedActivitiesinMedicalContextScientificDatabases

447

the palpation and the correct identiﬁcation of anterior

cervical anatomy and has the ability to record the con-

tact location of instruments on the trachea model dur-

ing the full duration of the simulated procedure.

2.1 Simulator Design

The trachea model is disposable and is replaced af-

ter each procedure. To minimize costs, the trachea

is made of cardboard with ﬁxed size and dimension

according to the trachea of an average adult. Foam

strips were cut for the cartilaginous tracheal rings and

were attached on the trachea with appropriate spac-

ing. The thyroid and cricoid cartilages are permanent

parts made of ABS plastic, using a 3D printer. These

components are ﬁxed onto a wooden base, and ﬁrmly

support the trachea model. Conductive foils are used

as low-cost sensors to detect six critical landmarks

(identiﬁed by A-F letters) that providers might con-

tact during the procedure. The conductive foils cover

landmarks on the trachea model. Only one of these six

landmarks, the cricothyroid membrane itself, is the

correct area to contact and make an opening. Other

landmarks like the posterior tracheal wall and lateral

locations into the tracheoesphageal grooves should be

avoided during the procedure.

The model is ﬁtted with an Arduino Uno micro-

controller board based on the Atmel Atmega 328 mi-

croprocessor with a mounted 8x8 LED matrix-based

display for user interface capability. The microcon-

troller records the contact data of the instruments

(scalpel, tracheal hook, and hemostat) onto the six

conductive foils, as each instrument is wired. Dur-

ing the procedure, when a closed circuit is detected

between the instrument and a patch of foil, the event

is recorded and labeled with the time in milliseconds.

Breaking contact is similarly recorded. Traditional

matrix scanning techniques are used by the microcon-

troller to detect connections between the instruments

and the foil patches. A minimum time (20 millisec-

onds) between contacts was used to debounce the in-

puts to the microcontroller. The resulting data was

later low-pass ﬁltered with a cutoff frequency of 10

Hz in accordance with general human reaction time.

The simulator’s design is shown in Figure 1.

Moreover, the design was optimized for materials

that are low-cost, widely available and simple to as-

semble. The total cost of the simulator was less than

$50, which has a lower price compared to existing

commercial simulators.

(a) Trachea model is covered with in-

ner bicycle tube as human skin.

(b) Six different landmarks repre-

sented by conductive foils and tools

(scalpel, hook and forceps) are con-

nected to microcontroller for data

collection.

Figure 1: Low-cost cricothyrotomy simulator.

2.2 How to Use the Simulator

Medical doctors who want to use the simulator are

ﬁrstly forced to watch a video tutorial published by

the New England Journal of Medicine (James and

Pacheco-Fowler, 2008; White et al., 2012b). After

watching the instructional materials, they are allowed

to perform the procedure on the simulator following

the instructions below:

• Step 1: Palpate the cricothyroid membrane. Im-

mobilize the larynx with the non-dominant hand

and perform the procedure with the dominant

hand.

• Step 2: Incise the skin (bicycle inner tube) verti-

cally after palpating the cricothyroid membrane.

• Step 3: Incise the cricothyroid membrane on tra-

chea model horizontally (1 cm length).

• Step 4: Insert the tracheal hook into cricoid carti-

lage.

• Step 5: Insert the hemostat and to expand the air-

way opening vertically and horizontally.

• Step 6: Insert the endotracheal tube.

As these steps were performed, all the data (con-

tact locations on trachea model, instruments informa-

tion, contact duration and total time) were recorded

DATA2014-3rdInternationalConferenceonDataManagementTechnologiesandApplications

448

by the microcontroller; procedures were also video-

recorded for analysis.

Thus, after the description of the used cricothy-

rotomy simulator, the importance of temporal infor-

mation is deﬁnitely clear. The following sections de-

scribe how our prototype has been designed and de-

veloped for helping medical doctors when they are us-

ing the simulator.

3 MODELING EXPECTED

ACTIVITIES

This section describes the model that we have deﬁned

in order to derive a formal deﬁnition of Expected Ac-

tivity for medical context. We use the temporal proba-

bilistic graph proposed by (Albanese et al., 2014; Al-

banese et al., 2011), so that the elapsed time between

observations also plays a role in deﬁning whether a

sequence of observations constitutes an activity, dif-

ferently from what happens in other models, such as

Hidden Markov Chains. We assume the existence of a

ﬁnite set S of action symbols, corresponding to atomic

events that can be detected by the Arduino microcon-

troller board, as described in Section 2.

3.1 Basic Deﬁnitions

An Expected Activity is a labeled directed graph A=

(V ,E, δ, ρ) where: (i) V is a ﬁnite set of nodes la-

beled with action symbols from S; (ii) E ⊆ V × V is

a set of edges; (iii) δ : E → N

associates with each

edge hv

i an upper bound of time that can elapse

between v

and v

; (iv) ρ : E → (0,1) is a function that

associates a probability distribution with the outgoing

edges of each node, i.e. ∀v ∈ V

∑

hv,v

i∈E

δ(hv, v

i) = 1;

(v) there exists an initial node I in the activity deﬁni-

tion, i.e. {v ∈ V | @ v

∈ V s.t. hv

,vi ∈ E} 6=

0 ; (vi)

there exists a ﬁnal node F in the activity deﬁnition,

i.e. {v ∈ V | @ v

∈ V s.t. hv,v

i ∈ E} 6=

We assume the existence of a ﬁnite set S of ac-

tion symbols representing particular interactions (for

instance, Pad C touched with SCALPEL, PAD C re-

leased) between medical doctors and the simulator.

Figure 2 shows an expected activity model represent-

ing simple interactions between a medical doctor and

the cricothyrotomy simulator.

Then, we deﬁne an instance of an expected activ-

ity as a speciﬁc path in A from the initial node to the

end node.

An instance of an Expected Activity (V, E, δ, ρ) is

a ﬁnite sequence hv

,...,v

i of nodes in V such that:

(i) hv

i+1

i ∈ E for 1 < i < m; (ii) {v | hv,v

i ∈ E} =

0, i.e. v

is the start node I; (iii) {v | hv

,vi ∈ E} = 0,

Figure 2: An example of Expected Activity Model.

Figure 3: System Architecture.

i.e. v

is the ﬁnal node F. The probability of the

instance is

∏

m−1

i=1

ρ(hv

i+1

We work with sequences of time-stamped events.

Let us assume that the number of observable events in

our domain is ﬁnite, each event can then be associated

to a different action symbol in the set S. We deﬁne an

observed event as a pair ω = (s,ts), where ω.s is the

action symbol associated to the event and ω.ts is the

time stamp at which s was observed.

We call a Simulator Log Ω a ﬁnite sequence of log

entries ω

Now, we are in the position of deﬁning the concept

of Activity Occurrence.

Let Ω be a Simulator Log and A=(V, E, δ, ρ) an

Expected Activity. An occurrence of A in Ω is a se-

quence h(ω

)...(ω

)i where: (i) hω

,...,ω

i is

a subsequence of Ω such as ω

= (ω

.ts,ω

.s), ω.s

being an action symbol from S and ω.ts the associ-

ated time-stamp; (ii) hv

,...,v

i is an instance of A;

(iii) v

= ω

.s for 1 < i < m

; (iv) ω

i+1

.ts − ω

.ts ≤

refers both to the node v

in A and the action symbol

labeling it

DiscoveringExpectedActivitiesinMedicalContextScientificDatabases

449

Figure 4: Expected Activity Models.

δ(hv

i+1

i) for 1 < i < m.

The probability p(o) of the occurrence o should be

the probability of the instance hv

,...,v

i. Of course,

shorter activities usually have higher probabilities.

Therefore, since we compare occurrences of different

activity models despite their different lengths, we in-

troduce the relative probability p

∗

(o) = p(o)/p(max).

When computing p

∗

(o) for a given occurrence o, we

consider p(max) as the highest probability of any in-

stance of A when ignoring each instance’s self-loops.

Thus, once we have given the previous formal

deﬁnitions for deﬁning our expected activity model,

we can describe the proposed architecture for ﬁnd-

ing expected activities in medical context scientiﬁc

databases in section 4.

4 THE PROPOSED

ARCHITECTURE

The theoretical model has been exploited to develop

a framework for the detection of expected activities

in medical context scientiﬁc databases. The structure

of the system is based on a modular architecture, as

shown in Figure 3, which allows the medical doctors

to get a very fast feedback about their performances

when using the simulator.

The following subsections describe the single

components of the overall system architecture.

4.1 The Arduino Microcontroller Board

As also mentioned in section 2, the Arduino micro-

controller board allows us to capture in real time

the contact data of the instruments (scalpel, tracheal

hook, and hemostat) from six different landmarks of

the simulator. In such a way, this component records

the series of time-stamped events, corresponding to

the medical doctors’ interactions with the simulator.

In more detail, events are deﬁned as the start and end

times of contacts between speciﬁc instruments and

surfaces on the anatomical model. Other types of

events are deﬁned in terms of readings from different

sensor types. Thus, events are represented by a se-

ries of symbols (ASCII characters). An excerpt of the

captured data is shown in Figure 5. Data are encoded

as follows:

• The ﬁrst single digit number indicates the instru-

ment (1 means Scalpel, 2 Hemostat and 3 Tra-

cheal Hook).

• The character indicates which foil patch is

touched: upper-case for making contact and

lower-case for breaking contact. In more detail, A

means Posterior tracheal wall, B the Right lateral

trachea and cricothyroid membrane, C the Mid-

line cricothyroid membrane (correct placement of

incision), D the Left lateral trachea and cricothy-

roid membrane, E the Cricoid cartilage and F the

Cartilaginous ring of lower tracheal wall.

• The last number is the time in milliseconds.

Then, the data captured in this way represent the

input of the Data Acquisition component.

4.2 The Data Acquisition Component

The Data Acquisition component includes an Adapta-

tion Module that converts the data captured using the

Arduino in a format suitable to the detection frame-

work (i.e. the Simulator Log): it also saves them into

a scientiﬁc database, which is able to store personal

DATA2014-3rdInternationalConferenceonDataManagementTechnologiesandApplications

450

information about the medical doctors who are using

the simulator as well.

Figure 5: Data captured using the Arduino Microcontroller

Board.

4.3 The Activity Detection Engine

The Activity Detection Engine takes as inputs time-

stamped user data collected in the Simulator Log and

a set of activity models to ﬁnd the activity occurrences

matching the known models. Such models have been

previously deﬁned by domain experts who have clas-

siﬁed them in two different categories: the good activ-

ities, corresponding to a correct use of the simulator

and the bad activities, corresponding to an incorrect

use of the simulator. Figure 4 shows two model ex-

amples of a good activity (at left), corresponding to

an excellent performance of the medical doctor and

a bad activity (at right), corresponding to a very bad

performance.

Expected activity occurrences in a data stream

are efﬁciently detected using tMagic (Albanese et al.,

2013), which allows to solve the problem of ﬁnd-

ing occurrences of high-level activity model in an ob-

served data stream. As a matter of fact, they propose

a data structure called temporal multiactivity graph to

store multiple activities that need to be concurrently

monitored, corresponding to our knowledge base of

good and bad activities. They then deﬁne an index

called Temporal Multiactivity Graph Index Creation

(tMAGIC) that, based on this data structure, examines

and links observations as they occur. Finally, they de-

ﬁne an algorithm to solve the evidence problem that

tries to ﬁnd all occurrences of an activity (with proba-

bility over a threshold) within a given sequence of ob-

servations. In this way, we are able to ﬁnd all the Ex-

pected Activities matching the activity models of our

knowledge base in a certain sequence of observations,

allowing medical doctors to get a very fast feedback

about their performances when using the cricothyro-

tomy simulator.

Figure 6: Framework Running Times.

5 PRELIMINARY

EXPERIMENTAL RESULTS

This section shows a preliminary experimental evalu-

ation of our framework. We present the experimental

protocol for evaluating our framework in terms of ex-

ecution time scalability, detection accuracy and user

satisfaction.

5.1 Evaluating Execution Time

We decided to measure

the execution time of our

framework for detecting Expected Activities in the

worst case (at least one action symbol for each mil-

lisecond) when varying the length of the Simulator

Log and using the previously deﬁned set of known ac-

tivity models. In more detail, the maximal length of

Simulator Log considered has been 5 minutes, since a

longer procedure would cause the death of the patient.

The time for acquiring data using the Arduino micro-

controller can be considered as negligible. Thus, the

Total Running Time is given by the sum of the Storing

and Index Building Time (higher value) and the Query

Time (lower value), as shown in Figure 6. However,

the obtained Total Running Time can be considered

low even if we are considering the worst case.

5.2 Accuracy Results

100 medical doctors participated in a trial and used

the simulator with our additional framework. The

classic Precision and Recall metrics have been used

All experiments presented in this Section were con-

ducted on a machine running Mac OS X 10.9.1, and mount-

ing a 2GHz Intel Core i7 processor with a 8 GB, 1600 MHz

DDR3.

DiscoveringExpectedActivitiesinMedicalContextScientificDatabases

451

Figure 7: User Satisfaction.

to compute the accuracy (Albanese et al., 2014; Al-

banese et al., 2011), by comparing the Expected Ac-

tivities discovered by our framework with a ground

truth deﬁned by experts who watched the recordings

of the medical doctors’performances several times.

We use {A

}

i∈[1,m]

to denote the Expected Activi-

ties returned by our framework and {A

}

j∈[1,n]

to de-

note the activities ﬂagged as expected by human an-

notators. Precision and recall were computed as fol-

lows:

P =

|{A

|∃A

s.t. A

≈ A

(1)

and

R =

|{A

|∃A

s.t. A

≈ A

(2)

We achieved an average Precision of 81% and an

average Recall of 98%, that can be considered a very

encouraging result.

5.3 User Satisfaction

After completing the procedure, the medical doc-

tors, classiﬁed in three different categories (expert,

medium-expert and not expert users), ﬁlled out a

questionnaire to report their level of training, expe-

rience and their impressions of the simulator. Each

subject was asked to answer 5 questions about the

simulator (How realistic, How anatomically accurate,

How educational, How useful, How real-time) using

a 5-point Likert scale. As we can see in Figure 7, sub-

jects (especially not-expert users) expressed positive

opinions about their experiences with the simulator.

6 CONCLUSIONS AND FUTURE

WORK

This work presented a framework for activity detec-

tion in the medical context. We started acquiring data

from a cricothyrotomy simulator, when used by medi-

cal doctors and we then stored the captured data into a

scientiﬁc database. Finally, we used some stable ac-

tivity detection algorithms for discovering expected

activities, corresponding to speciﬁc performances ob-

tained by the medical doctors when using the simu-

lator. Some preliminary experiments showed encour-

aging results concerning efﬁciency, effectiveness and

user satisfaction.

Future work will be devoted to enlarge our exper-

imentation and to plan to integrate the prototype in

more complex and thorny applications by adding new

functionalities and, if necessary, additional layers to

the overall system architecture. For example, a po-

tential application of this tool could consist in mon-

itoring the performance of medical personnel in real

time and detecting potential safety hazard in advance,

for instance, using machine learning techniques and

observations learned during the training of medical

personnel. Moreover, data mining techniques could

be used in an ofﬂine setting to analyze in detail the

medical doctors’performances.

REFERENCES

Albanese, M., Molinaro, C., Persia, F., Picariello, A., and

Subrahmanian, V. (2014). Discovering the top-k unex-

plained sequences in time-stamped observation data.

IEEE Transactions on Knowledge and Data Engineer-

ing, 26(3):577–594.

Albanese, M., Molinaro, C., Persia, F., Picariello, A., and

Subrahmanian, V. S. (2011). Finding ”unexplained”

activities in video. In Proceedings of the Twenty-

Second International Joint Conference on Artiﬁcial

Intelligence - Volume Volume Two, IJCAI’11, pages

1628–1634. AAAI Press.

Albanese, M., Moscato, V., Picariello, A., Subrahmanian,

V. S., and Udrea, O. (2007). Detecting stochastically

scheduled activities in video. In Veloso, M. M., editor,

IJCAI, pages 1802–1807.

Albanese, M., Pugliese, A., and Subrahmanian, V. S.

(2013). Fast activity detection: Indexing for tempo-

ral stochastic automaton-based activity models. IEEE

Trans. Knowl. Data Eng., 25(2):360–373.

Bennett, B., Cailteux-Zevallos, B., and Kotora, J.

(2011). Cricothyroidotomy bottom-up training re-

view: battleﬁeld lessons learned. Military Medicine,

176(11):1311–9.

Boselli, R., Cesarini, M., Mercorio, F., and Mezzanzanica,

M. (2014a). Planning meets data cleansing. In Pro-

DATA2014-3rdInternationalConferenceonDataManagementTechnologiesandApplications

452

ceedings of the Twenty-Fourth International Confer-

ence on Automated Planning and Scheduling.

Boselli, R., Cesarini, M., Mercorio, F., and Mezzanzan-

ica, M. (2014b). A policy-based cleansing and in-

tegration framework for labour and healthcare data.

In Holzinger, A. and Jurisica, I., editors, Interactive

Knowledge Discovery and Data Mining in Biomedical

Informatics, volume 8401 of Lecture Notes in Com-

puter Science, pages 141–168. Springer Berlin Hei-

delberg.

Brand, M., Oliver, N., and Pentland, A. (1997). Coupled

hidden markov models for complex action recogni-

tion. In Proceedings of the 1997 Conference on Com-

puter Vision and Pattern Recognition (CVPR ’97),

CVPR ’97, pages 994–999, Washington, DC, USA.

IEEE Computer Society.

Combi, C., Montanari, A., and Pozzi, G. (2007). The t4sql

temporal query language. In Silva, M. J., Laender, A.

H. F., Baeza-Yates, R. A., McGuinness, D. L., Olstad,

B., Olsen, . H., and Falco, A. O., editors, CIKM, pages

193–202. ACM.

Cook, T., Woodall, N., and Frerk, C. (2011). Major com-

plications of airway management in the uk: results

of the fourth national audit project of the royal col-

lege of anaesthetists and the difﬁcult airway society.

part 1: anaesthesia. Fourth National Audit Project,

106(5):617–31.

Cuntoor, N., Yegnanarayana, B., and Chellappa, R. (2008).

Activity modeling using event probability sequences.

Image Processing, IEEE Transactions on, 17(4):594–

607.

Hamid, R., Huang, Y., and Essa, I. (2003). Argmode - ac-

tivity recognition using graphical models. In Com-

puter Vision and Pattern Recognition Workshop, 2003.

CVPRW ’03. Conference on, volume 4, pages 38–38.

Hongeng, S. and Nevatia, R. (2001). Multi-agent event

recognition. In Computer Vision, 2001. ICCV 2001.

Proceedings. Eighth IEEE International Conference

on, volume 2, pages 84–91 vol.2.

James, H. and Pacheco-Fowler, V. (2008). Cricothyrotomy.

N Engl J Med.

King, D., Ogilvie, M., Michailidou, M., Velmahos, G.,

Alam, H., deMoya, M., and Fikry, K. (2012). Fifty-

four emergent cricothyroidotomies: are surgeons re-

luctant teachers? Scand J Surg, 101(1):13–5.

Oliver, N., Horvitz, E., and Garg, A. (2002). Layered rep-

resentations for human activity recognition. In Mul-

timodal Interfaces, 2002. Proceedings. Fourth IEEE

International Conference on, pages 3–8.

Patrick, J.D. Safari, L. and Cheng.Y. (2013). Knowledge

discovery and knowledge reuse in clinical information

systems. Proc. The 10th IASTED International Con-

ference on Biomedical Engineering (BioMed 2013).

Safari, L. and Patrick, J. (2013). A temporal model for

clinical data analytics language. In Engineering in

Medicine and Biology Society (EMBC), 2013 35th

Annual International Conference of the IEEE, pages

3218–3221.

Vaswani, N., Roy-Chowdhury, A., and Chellappa, R.

(2005). ”shape activity”: a continuous-state hmm for

moving/deforming shapes with application to abnor-

mal activity detection. Image Processing, IEEE Trans-

actions on, 14(10):1603–1616.

Wang, H., Mann, N., Mears, G., Jacobson, K., and Yealy,

D. (2011). Out-of-hospital airway management in the

united states. Resuscitation, 82(4):378–385.

Warner, K., Sharar, S., Copass, M., and Bulger, E. (2009).

Prehospital management of the difﬁcult airway: a

prospective cohort study. J Emerg Med, 36(3):257–

65.

White, L., Bly, R., D’Auria, D., Aghdasi, N., Bartell, P.,

Cheng, L., and Hannaford, B. (2013). Cricothyro-

tomy simulator with computational skill assessment

for procedural skill training in the developing world.

In AAO-HNSF Annual Meeting and OTO Expo.

White, L., Bly, R., D’Auria, D., Aghdasi, N., Bartell, P.,

Cheng, L., and Hannaford, B. (2014). Cricothyrotomy

simulator with computational skill assessment for pro-

cedural skill training in the developing world. Journal

of Otolaryngology - Head and Neck Surgery.

White, L., Bly, R., D’Auria, D., Bly, R., Aghdasi, N.,

Bartell, P., Jones, C., and Hannaford, B. (2012a).

Cricothyrotomy simulator training for the developing

word. In 2012 IEEE Global Humanitarian Technol-

ogy.

White, L., Bly, R., D’Auria, D., Bly, R., Aghdasi, N.,

Bartell, P., Jones, C., and Hannaford, B. (2012b).

Global simulation training in healthcare.

Zhou, L. and Hripcsak, G. (2007). Methodological review:

Temporal reasoning with medical data-a review with

emphasis on medical natural language processing. J.

of Biomedical Informatics, 40(2):183–202.

DiscoveringExpectedActivitiesinMedicalContextScientificDatabases

453