Ayman Hajja

Reduction of Readmissions to Hospitals Based on Actionable Knowledge Discovery and Personalization

In this work, we define procedure paths as the sequence of procedures that a given patient undertakes to reach a desired treatment. In addition to its value as a mean to inform the patient of his or her course of treatment, being able to identify and anticipate procedure paths for new patients is an essential task for examining and evaluating the entire course of treatments in advance, and ultimately rectifying undesired procedure paths accordingly. In this paper, we first introduce two approaches for anticipating the state of the patient that he or she will end up in after performing some procedure p; the state of the patient will consequently indicate the following procedure that the patient is most likely to undergo. By clustering patients into subgroups that exhibit similar properties, we improve the predictability of their procedure paths, which we evaluate by calculating the entropy to measure the level of predictability of following procedure. The clustering approach used is essentially a way of personalizing patients according to their properties. The approach used in this work is entirely novel and was designed specifically to address the twofold problem of first being able to predict following procedures for new patients with high accuracy, and secondly being able to construct such groupings in a way that allows us to identify exactly what it means to transition from one cluster to another. Then, we further devise a metric system that will evaluate the level of desirability for procedures along procedure paths, which we would subsequently map to a metric system for the extracted clusters. This will allow us to find desired transitions between patients in clusters, which would result in reducing the number of anticipated readmissions for new patients.

Hierarchical object-driven action rules

In this work, we present the hierarchical object-driven action rules; a hybrid action rule extraction approach that combines key elements from both the classical action rule mining approach, first proposed by Raś and Wieczorkowska (2000), and the more recent object-driven action rule extraction approach proposed by Hajja et al. (2012, 2013), to extract action rules from object-driven information systems. Action rules, as defined in Raś and Wieczorkowska (2000), are actionable tasks that describe possible transitions of instances from one state to another with respect to a distinguished attribute, called the decision attribute. Recently, a new specialized case of action rules, namely object-driven action rules, has been introduced by Hajja et al. (2012, 2013). Object-driven action rules are action rules that are extracted from information systems with temporal and object-based nature. By object-driven information systems, we mean systems that contain multiple observations for each object, in which objects are determined by an attribute that assumingly defines some unique distribution; and by temporally-based information systems, we refer to systems in which each instance is attached to a timestamp that, by definition, must have an intrinsic meaning for each corresponding instance. Though the notion of object-driven and temporal-based action rules had its own successes, some argue that the essence of object-driven assumptions, which is in big part the reason for its effectiveness, are imposing few limitations as well. Object-driven approaches treat entire systems as multi-subsystems for which action rules are extracted from; as a result, more accurate and specific action rules are extracted. However, by doing so, our diverseness of the extracted action rules are much less apparent, compared to the outcome when applying the classical action rule extraction approach, which treats information systems as a whole. For that reason, we propose a hybrid approach which builds a hierarchy of clusters of subsystems; a novel way of clustering through treatments responses similarities is introduced.

Gamesourcing to acquire labeled human pose estimation data

In this paper, we present a gamesourcing method for automatically and rapidly acquiring labeled images of human poses to obtain ground truth data as input for human pose estimation from 2D images. Typically, these datasets are constructed manually through a tedious process of clicking on joint locations in images. By using a low-cost RGBD sensor, we capture synchronized, registered images, depth maps, and skeletons of users playing a movement-based game and automatically filter the data to keep a subset of unique poses. Using a recently-developed, learning-based human pose estimation method, we demonstrate how data collected in this manner is as suitable for use as training data as existing, manually-constructed data sets.

Pair-Based object-driven action rules

Action rules, as proposed by Raś and Wieczorkowska in [11], can be defined as actionable tasks that describe possible transitions of objects from one state to another with respect to a distinguished attribute. Recently, a new specialized case of action rules, namely object-driven action rules, has been introduced by Ayman et al. in [4]. Object-driven action rules are action rules that are extracted from information systems with temporal and object-based nature. By object-based nature, we refer to systems that contain multiple observations for each object. A typical example of an object-based system would be a system of patients recording multiple visits; each patient is considered a distinct object. In this paper, we will further investigate the concept of object-driven action rules by proposing a new pair-based way of examining object-driven systems, which we believe is more intuitive for temporal and object-driven systems. The focus of this paper will be on our proposed pair-based approach, along with the modifications required to extract action rules and calculate their properties.

Predicting the Primary Medical Procedure Through Clustering of Patients’ Diagnoses

Healthcare spending has been increasing in the last few decades. This increase can be attributed to hospital readmissions, which is defined as a re-hospitalization of a patient after being discharged from a hospital within a short period of time. The correct selection of the primary medical procedure by physicians is the first step in the patient treatment process and is considered to be of the main causes for hospital readmissions. In this paper, we propose a recommender system that can accurately predict the primary medical procedure for a new admitted patient, given his or her set of diagnoses. The core of the recommender system relies on identifying other existing patients that are considered similar to the new patient. That said, we propose three approaches to predict the primary procedure. The results show the ability of our proposed system to identify the primary procedure. It can be later used to build a graph which shows all possible paths that a patient may undertake during the course of treatment.

Multipurpose Web-Platform for Labeling Audio Segments Efficiently and Effectively

One of the principal reasons for the success of machine learning discoveries can be attributed to the utilization of large sums of labeled datasets used to train various learning models. The availabilities of annotated data depend, to a large extent, on the nature of the domain, and how easy it is to obtain labeled data-points. One of the areas that we believe still lacks substantial labeled data is audio. This is not surprising, since labeling audio segments can be rather tedious and time-consuming, mainly due to the temporal nature of it. In this paper, we present a free and open-source web-based platform that we developed, which allows individuals and research teams to crowdsource large sums of labeled audio segments efficiently and effectively. Once an individual or a team signs up to use the platform as researchers, they will be granted administrative access that will enable them to upload their own audio files, and customize the labeling and data collection process according to their study needs. Examples of customizing the study include listing the different labels of interest, specifying the duration of audio segments and how they should be extracted from the audio file(s), and dictating how labelers should be prompted with the audio segments based on a set of pre-determined user-defined rules. Our system will automatically handle generating the audio segments from the audio files, presenting labelers with an intuitive interface using the rules specified by the study administrators, and finally recording the labelers’ responses and providing them to the administrators of the study in a readable and easy-to-access format.

Segment-Removal Based Stuttered Speech Remediation

Speech remediation by identifying those segments which take away from the substance of the speech content can be performed by correctly identifying portions of speech which can be deleted without diminishing from the speech quality, but rather improving the speech. Speech remediation is especially important when the speech is disfluent as in the case of stuttered speech. In this paper, we describe a stuttered speech remediation approach based on the identification of those segments of speech which when removed would enhance speech understandability in terms of both speech content and speech flow. The approach we adopted consists of first identifying and extracting speech segments that have weak significance due to their low relative intensity, then classifying the segments that should be removed. We trained several classifiers using a large set of inherent and derived features extracted from the audio segments for the purpose of automatic improvement of stuttered speech by providing a second layer filtering stage. This second layer would discern the audio segments that need to be eliminated from the ones that do not. The resulting speech is then compared to the manually-labeled “gold standard” optimal speech. The quality comparisons of the resulting enhanced speeches and their manually-labeled counterparts were favorable and the corresponding tabulated results are presented below. To further enhance the quality of the classifiers we adopted a voting techniques that encompassed an extended set of models from 14 algorithms and presented the classifier performance measures from different voting threshold values. This voting approach allowed us to improve the specificity of the classification by reducing the false positive classifications at the expense on additional false negatives thus improving the practical effectiveness of the system.