Clinical record integration and visualization is one of the most important abilities of modern health information systems (HIS). Its use on clinical encounters plays a relevant role in the e cacy and efciency of healthcare. However, integrated HIS of central hospitals may gather millions of clinical reports (e.g. radiology, lab results, etc.). Hence, the clinical record must manage a stream of reports being produced in the entire hospital. Moreover, not all documents from a patient are relevant for a given encounter, and therefore not visualized during that encounter. Thus, the HIS must also manage a stream of events of visualization of reports, which runs in parallel to the stream of documents production. The aim of our project is to provide the physician with a recommendation of clinical reports to consider when they log in the computer. Our approach is to model relevance as the probability that a given document will be accessed in the current time frame. For that, we design a data stream management system to process the two streams, and Bayesian networks to learn those probabilities based on document, patient, department and user information. One of the biggest challenges to the learning problem, so far, is that no negative examples are produced by the stream (i.e. there are no record of documents not being visualized) leading to a one-class classi cation problem. The aim of this paper is to clearly present the setting and rationale for the approach. Current work is focused on both the stream processing mechanism and the Bayesian probability estimation.
The identi cation of clinically relevant information should enable an improvement in user interface design and in data management. However, it is di cult to identify what information is important in daily clinical care, and what is used only occasionally. The main problem addressed by this project is how to estimate the relevance of healthcare information in order to anticipate its usefulness at a speci c point of care. In particular, we want to estimate the probability of a piece of information being accessed during a certain time interval, taking into account the type of data, the context where it was generated and is needed and the type of users who access it, and to use this probability to prioritize the information. 1.1
In the healthcare domain, and especially in critical and acute care, the age of data
is one of the factors often used to assess data relevance, making new information
more relevant to the current search. Some authors have categorized old data as
data at least three days old. However, in a previous study, the authors have
examined for how long are clinical documents used by health professionals in
a hospital environment, and how this is associated with document content and
the context of information request [
The current setting is a central hospital that has several departmental
information systems that produce clinical documents (e.g. radiology reports and lab
results) that might be relevant for the practice of healthcare. The access to this
documents is better achieved by a centralized information system that integrates
all the di erent departmental systems, aggregating the documents that are most
relevant for the current encounter [
The rst stream is gathered by integrating documents from heterogeneous information systems, so we might end up with syncronization issues. In the current setting, the central hospital information system receives an increasing rate of 200+ documents per hour (as seen in Figure 1), relative to the daily 5300+ patients in the hospital (including inpatients, outpatients and emergency rooms) which need to be processed. This created a pool of 8M+ documents (produced since 2004 from 400K+ patients). However, due to constant document revisions, \only" 2.9M+ are actually available for visualization (active documents), whilst 5M+ are previous versions. 0 5 2 0 0 2 0 5 /rsouH 150 t n e couDm 100
● VPirsoudaulcizeadtioDnoscuinm2e0n1ts0/2011 0 ● 2004 ● 2005 ● 2006 ● 2007 Year ● 2008 ● 2009 ● 2010
The second stream is controlled in the centralized integrating information system, tracking information on visualizations performed by 4850+ users. Older documents tend to be less visualized in encounters but, as seen in Figure 1, nearly half of the visualizations in 2010 and 2011 targeted documents older than January 2009, so old documents cannot be discarded. During an encounter with a patient, the pool of documents that one of the currently active 2375 users might access is 537.9 (average number of active documents per active patient). Clearly, the user cannot be presented with a list of 530+ documents, so a ranking is needed to prioritize the most relevant. 1.3
Overall, this is one clear setting of data streams in medical scenarios, on which
machine learning techniques can be applied to improve healthcare. In order to
select relevant documents for visualization in a precise encounter, we need to
take care of both the age and the information related to those documents.
Either way, this setting is an uncertain one. Thus, the best way to model this
relevance is to estimate the probability that the document is going to be
visualized in the current time period. In machine learning, uncertainty is usally well
modeled using Bayesian approaches [
The aim of this work is to extend the health information system responsible for visualization of clinical documents at point of care, providing the physician with a recommendation of clinical reports to consider when in the presence of a patient. Basically, we propose to study how to estimate the relevance of healthcare information in a particular setting, aiming to use this to create adaptive user interfaces. Speci cally, our objectives are: { to collect and prepare log data from hospital information systems usage, to feed our two data streams; { to study the factors associated with the relevance of clinical documents; { to de ne an algorithm to estimate the relevance of a particular document for a precise encounter; { to implement an adaptive user interface based on a ranked list of recommended documents per encounter.
The paper is organized as follows. Next section presents background knowledge on electronic health records, learning from data streams and Bayesian networks in healthcare. Then, section 3 presents our approach for i) the data stream processing, ii) the estimation strategy, iii) the incremental learning and iv) the recommendation generator. Section 4 ends the exposition with future work. 2
Background
This work is related with three di erent areas of research: medical
informatics, especially devoted to electronic health records; Bayesian learning from data
streams; and one-class classi cation (check [
The practice of medicine has been described as being dominated by how well
information is collected, processed, retrieved, and communicated [
Currently in most hospitals there are great quantities of stored digital data
regarding patients, in administrative, clinical, lab or imaging systems. An
important challenge is to guarantee the optimal conditions for health professionals to
access clinical data while hospital information systems are still being developed.
Although great advances have been made over the years, on-demand access to
clinical information is still inadequate in many settings, contributing to
duplication of e ort, excess costs, adverse events, and reduced e ciency. Although it is
widely accepted that full access to integrated electronic health records (EHRs)
and instant access to up-to-date medical knowledge signi cantly reduces faulty
decision making resulting from lack of information [
Shapiro et al. found that, although emergency department doctors believe
their patients would bene t from longitudinal records, they only try to obtain
such data in 10% of the cases [
Cruz-Correia et al. have done several pilot studies to analyse for how long
clinical documents are useful for health professionals in a hospital environment,
bearing in mind document content and the context of the information request.
The results show that some clinical reports are still used one year after creation,
regardless of the context in which they were created, although signi cant di
erences existed in reports created during distinct encounter types. The median-life
of reports by the type of encounter during which they were created is 1.7 days
for emergency, 3.9 days for inpatient and 27.7 days for outpatient encounters.
They conclude that the usage of patients past information (data from previous
hospital encounters), varied signi cantly according to the setting of healthcare
and content [
Also, the amount of digital data produced in the medical imaging departments has increased rapidly in recent years, due mainly to two factors: 1) greater use of additional diagnostic procedures, resulting in a greater number of tests produced and, 2) an increase in the quality of the examinations, which is translated into a greater number of images acquired. Despite great development in digital storage technologies, the cost and e ort required to maintain this information online throughout its life cycle can be considerable.
The management of information in these systems is usually implemented using Hierarchical Storage Management (HSM) solutions. This type of solution enables the implementation of various layers which use di erent technologies with di erent speeds of access, corresponding to di erent associated costs. However, the solutions which are currently implemented in Picture Archiving and Communication System (PACS) use simple rules for information management, based on variables such as the time elapsed since the last access or the date of creation of information, not taking into account the likely relevance of information in the clinical environment.
Classifying the relevance of information based only on the time elapsed since the date of acquisition is clearly ine cient. It is expected that the need to consult an examination at a given time will be dependent on several factors beyond the date of the examination, such as type of examination and the patient's pathology. Thus, a system that uses more factors to identify the relevance of information at a given time would be more e cient in managing the information that is stored in fast memory and slow memory. 2.2
What distinguishes current data from earlier one are automatic data feeds. We do
not just have people who are entering information into a computer. Instead, we
have computers entering data into each other [
A data stream is an ordered sequence of instances that can be read only once
or a small number of times using limited computing and storage capabilities.
The data elements in the stream arrive online, being potentially unbounded in
size. Once an element from a data stream has been processed it is discarded or
archived. It cannot be retrieved easily unless it is explicitly stored in memory,
which is small relative to the size of the data streams. These sources of data are
characterized by being open-ended, owing at high-speed, and generated by non
stationary distributions [
In online streaming scenarios, predictions are usually followed by the real
label value in a short future (e.g., prediction of next value of a time series).
Nevertheless, there are also scenarios where the real label value is only available
after a long term, such as predicting one week ahead electrical power
consumption [
The sequences of data points are not independent, and are not generated
by stationary distributions. We need dynamic models that evolve over time and
are able to adapt to changes in the distribution generating examples [
Hulten et al. [
The application of data mining and machine learning techniques to medical
knowledge discovery tasks is now a growing research area. These techniques vary
widely and are based on data-driven conceptualizations, model-based de nitions
or on a combination of data-based knowledge with human-expert knowledge [
The de nition of clinical decision support systems is now a major topic since
it may help the diagnosis, the prognosis of rate of mortality, the prognosis of
quality of life, or even treatment selection. However, the complicated nature of
real-world biomedical data has made it necessary to look beyond traditional
biostatistics [
Traditional statistical methods require that the model structure is given and
only probabilistic information is learned from biomedical evidence, in the form
of data, whereas machine-learning approaches enable that both the structure of
the models and the probabilistic information are evidence-based[
Given their improved management of uncertainties, Bayesian networks have
been successfully applied in healthcare domains [
On a general basis, a Bayesian network represents a joint distribution of one
set of variables, specifying the assumption of independence between them, with
the inter-dependence between variables being represented by a directed acyclic
graph. Each variable is represented by a node in the graph, and is dependent of
the set of variables represented by its ascendant nodes; a node X is a ascendant
of another node Y if exists a direct arc from X to Y [
After the qualitative and quantitative models are constructed, the next step,
and one of the most important, is how to calculate the new probabilities when
new evidence is introduced in the network. This process is called inference and
works as follows. Each variable has a nite number of categories greater than
or equal to two. A node is observed when there is knowledge about the state
of that variable. The observed variables have a huge importance because with
conditional probabilities they de ne the prior probabilities of the non observed
variables. With the joint probabilities we can calculate the marginal probabilities
of each unobserved variable, adding for all categories the probabilities that the
variable is in the desired state [
Optimizing the visualization of the clinical record This section exposes our approach to the problem, presenting the data stream management system, the relevance estimation strategy, the Bayesian learning model, and the documents recommender system. 3.1
Currently, a lot of patient information is accessible to healthcare professionals at the point of care. In some cases, the amount of information is becoming too large to be readily handled by humans or to be e ciently managed by traditional storage algorithms. Most Hospital information systems record the actions performed by users { the log le. These logs are kept for audit purposes but can give insights into the information needs of healthcare professionals in a particular situation. The study of these logs should allow us not only to describe how the systems were used, but may also be useful to predict future use of the system and of the data items it contains. To this latter task, there are two data streams being produced in parallel: the stream of documents being created, and the stream of visualization events.
The rst stream is gathered by integrating documents from heterogeneous
information systems, which should be modeled according to the insert-delete
or turnstile model [
By applying regression methods or other modeling techniques it is possible to
identify which factors are associated with the usage or relevance of patient data
items. These factors and associations can then be used to estimate data relevance
in a speci c future time interval. We expect to nd considerable di erences in
the relevance or median-life of information depending on several factors, such
as the origin of the data (e.g. lab department, emergency department), the age
of patient, and the current patient diagnosis. We also expect there to be an
exponential decay in the use and thus relevance of each patient's data over
time [
As previously discussed, we rely on the Bayesian networks ability to model
uncertainty to estimate the probability of a given clinical document to be
visualized. After de ning the set of patient variables P and the set of other factors F
that are associated with that relevance, we shall create a Bayesian network with
jjP jj + jjF jj + 2 nodes, where the remaining nodes are the age of the document
and the class (visualized or not). Given the discrete characteristics of variables in
the Bayesian network, age of document needs to be categorized into contiguous
intervals. According to previous work [
When a document is visualized (i.e. a new observation is created in the visualization stream) a learning example can be created, computing the age of the document with real creation and visualization dates. This learning example can thus be fed to the Bayesian network. However, there are no records of non-visualization of documents, so negative learning examples are never created. Sophisticated approaches shall be tested to solve this issue: 1. for each positive example representing a visualization event, create as many negative examples for the time periods elapsed since the last visualization (or creation); or 2. de ne a purely one-class classi cation learner.
From our setting, we can easily extract information on previous visualization of a given clinical document (the log le includes this data) so we will follow the rst approach. 3.4
The goal is to achieve the following setup. A doctor-patient encounter (e.g. emergency, outpatient consultation, inpatient consultation) requires visualization of clinical documents. These could be documents from the same patient (patient's medical history) or documents from patients with similar characteristics or diagnosis. Given the huge amount of available clinical documents, the information system should list only the most relevant for that encounter.
Several approaches to rank the recommendations could be followed, and certainly they will be studied. At rst, we shall consider the probability of a clinical document being visualized as the single rank variable. However, if we need to include documents from di erent patients, we might end-up biasing the results by considering erroneous visualizations. Naively, the probabilities of visualization of all possible documents need to be computed everytime the system needs to list the recommendations, which will turn the process infeasible given the streaming setup. There are, at least, three possible ways to solve this issue: 1. at every visualization event, probability of visualization of only that document is updated; 2. at every visualization event, probability of visualization of all documents is updated; 3. at every ranked list request, relevance of that patient's or similar patient's documents is updated;
This is the least solved part of the system, and future work is expected. Nevertheless, we feel that it is important to keep this target in mind because it may condition the possible paths that previous modules are going to traverse. 4
Future steps and expected impact of the system Future work is concentrated on: a) de ning a data stream management system, b) de ning the factors that in uence the relevance of clinical documents, c) build learning models to estimate the relevance of a single document for a given encounter, d) generate recommendations based on a ranking, and e) develop and test the prototype with real data. To our knowledge, the use of machine learning techniques to support graphical user interfaces and management storage systems in healthcare information systems is novel and could be an important contribution to science and likely to be incorporated into commercial products in the future. The results of this research are also likely to have an important impact on the quality of healthcare by further increasing the usability and intelligence of existing information systems.
This work is supported by Portuguese Foundation for Science and Technology (FCT) under projects OPTIM (PTDC/EIA-EIA/099920/2008) and KDUDS (PTDC/EIA-EIA/98355/2008).