Automated Assessment of Speech Development in Pre-school Children

In Australia, speech and language impairments in children aged 4 – 5 years is a high prevalence disorder. The incidence of these impairments in preschool children range from 15–22% and between 12–15% in school- age children. The social and economic impact of speech and language impairments is pervasive. A further negative outcome of speech and language impairment in preschool children is an increased risk of psychiatric disorders. Children with a history of language impairment are almost twice as likely to develop a psychiatric disorder by age 19, in contrast to those without a history of language impairment.

In this project we are seeking to develop a cost-effective and ubiquitous solution for early diagnosis of the language development problem. In particular, we aim to develop signal processing algorithms to extract information about the language development from the spontaneous speech and develop classifiers to automatically classify speech segments with language impairment in a natural setting at home.

Collaborators: Queensland Health, Centre for Clinical Research, University of Queensland and Royal Bris- bane Children Hospital.

Guiding Ebola Patients to Suitable Health Facilities: An SMS-based Approach

Difficult access due to remoteness and lack of transportation [10] were some of the main reasons that widened the gap between the patient and the healthcare provider during the last Ebola outbreak in West Africa. Access to information that guides patients to the nearest facility with appropriate resources was another challenge. There is currently no known system that can improve this access.

We proposed building a recommendation system based on simple SMS text messaging to help Ebola patients readily find the closest health service with available and appropriate resources. The system will map peo- ples reported symptoms to likely Ebola case definitions and suitable health service locations. This will be extremely valuable for heath workers to better plan and anticipate responses to the current Ebola Outbreak in West Africa and hopefully prevent many new cases.

Collaborators: Distributed Sensing Systems, CSIRO, Medicins Sans Frontieres.

Falls Risk Assessment System on Hearing Aid

The aim of the project is to develop hardware and complex signal processing algorithms, which can be integrated within the smaller form factor of a hearing aid and can be used to measure heart rate and detect falls. Using the proposed technology we seek to detect falls and assess its severity by continuously monitoring subject’s heart rate after the fall has occurred. In particular, the project seeks to embed a 3-axis accelerometer, gyroscope and a microphone in the hearing aid, where the microphone will be used to detect the heart rate and the 3-axis accelerometer and the gyroscope will be used jointly to detect abnormal body movement to infer falls. Our system is built upon two assumptions: First, when the earphone of the hearing aid is placed in the ear canal, heartbeat is the only periodic sound that the microphone should ideally gather [3]. Second, fall creates an explicit pattern in the accelerometer response, which is distinguishable from other activities [4]. Collaborators: Royal Brisbane Women’s Hospital Falls Risk Assessment Unit, and Blamey and Saunders Hears.

An Intelligent Sensing Framework and Alert System to Support Efficient Interventions to Reduce In- patient Falls

Conventional bed-exit alarm systems rely on simple sensors, which trigger upon a change in patient position (such as sitting up or coming up to standing). Recent advances in bed-exit alarm system have included improved sensors and multiple criteria for triggering but still lack the functionality to support graded and risk-appropriate responses. Our system will constitute a substantial advancement over current approaches in its non-intrusive nature, its potential to accurately reject situations that do not require intervention, and to recognize and therefore alert staff to specific high risk movement behaviors. We will achieve this by matching a highly sensitive data sampling platform with pattern recognition algorithms that scan the data in real-time for specific movement markers. The system would then reject safe bed-exit patterns, and in-bed activity and only trigger a smaller number of graded and information-rich alerts that allow staff to prioritize the type of response required. Furthermore, the open architecture of the solution will allow for software and sensor improvements for future expansion of functionality (such as monitoring applications in patients with delirium, sleep/activity tracking and pressure injury risk monitoring).

Collaborators: Royal Brisbane Women’s Hospital Falls Risk Assessment Unit.

24/7 Personalised Smart Carer for Treatment Monitoring and Early Interventions in Dementia

Dementia is a leading cause of disability and dependency among older people worldwide, with significant physical, psychological, social and economical impact on carers, families and society. Currently, there is no cure, nor ways to modify disease progression, with numerous treatments in various stages of investigation. Much can be, however, offered to support and improve the lives of people with dementia, and their carers and families. We believe smart devices hold great potential to improve dementia care. By combining wearable,

wireless and big data technologies, smart devices could continuously monitor the physiological, behavioral, and emotional states of the patients, and communicate them to their carers and physicians in real time. Such system will enable us, for the first time, to track disease progression and monitor treatment efficacy 24/7, and provide high quality data for developing personalized care plan for individual patient.

Collaborators: Queensland Institute of Medical Research. 

Compressive Sensing for Big Data Analytics

Big Data leverages tremendous potential to deeply understand the world around us, however, concurrently it poses extreme challenges of managing and making sense of the enormous data pool. Encouragingly, compres- sive sensing, which is an emerging signal processing technique, offers data reduction while acquisition. This is a breakthrough compared to other traditional transform coding methods - where data footprint is reduced after acquisition of complete data. Another promising aspect of compressive sensing is that many decisions (e.g., fault detection in Sensor Network [1]) can be made by directly using the encoded (or compressed) data; this could alleviate the big data problem to a great extent.

Automatic inference of Activities of Daily Living

In the high level, this project is targeted to assess independent living capacity of elderlies [13]. In current practice, the activities of daily living or ADLs are assessed by clinical personnel, which is not very cost- effective. In this project, we aim to develop an automated and cost-effective system to assess the ADL for individuals. This project imposes three attributes/constraints: non-wearable, context-aware and non-invasive. We compensate for invasive sensors (e.g. camera) using plethora of passive sensors, such as power meters, on-bed accelerometer, motion sensors, and contact sensors. We apply sensor fusion on the sensor feed to extract information about various activities of daily living, including, feeding, hygiene, dressing, transfer, mobility, social interaction, sleep efficiency and so on. In particular, using the multimodal sensor fusion we model baseline behavior for individuals and then develop algorithms to detect any persistent deviation subject to context information. Context information is important for anomaly detection. For example, a person might not attend to morning walk not because of physical illness but due to a morning shower. In this project we also analyze multi-modal sensor data to model the correlation between various physical behaviors with cognitive functions and develop methods for early detection of various neurodegenerative disorders.

Temporal-spatial Noise mapping from Crowdsourced Data

I started my PhD journey with this crowdsourced noise mapping application, which I named Ear-Phone [9]. The key fundamental problem in Ear-Phone is signal reconstruction from incomplete or insufficient crowd- sourced data. I took up Compressive Sensing (CS) as a tool to address the problem. Compressive sensing is a new paradigm of signal processing, which under sparsity constrains offers signal reconstruction from a small number of measurements. However, the signal reconstruction problem of Ear-Phone was not perfectly aligned with the compressive sensing framework. Compressive measurement needs to be taken on the complete signal, but crowdsourced data cannot guarantee the capture of the complete signal due to uncoordinated sampling. We showed that using additional measurements over the requirement of the compressive sensing theory, we can still accurately reconstruct a signal. This work has been featured by many prestigious science review portals including MIT Technology Review, WIRED, ACM Communications, ABC Science and many more.

Non-uniform Sampling in Rechargeable Wireless Sensor Networks

In this project, I investigated the feasibility of signal reconstruction at energy neutral condition from recharge- able wireless sensor nodes, where nodes possess non-uniform energy profile due to uneven solar energy scavenging opportunity in the forest [6]. Compressive sensing only supported uniform sampling; we made

theory contributions to enable non-uniform sampling via compressive sensing. The outcome from this work was a distributed non-uniform sampling algorithm built on a compressive sensing framework. Using similar sensor network datasets we further verified that the proposed non-uniform sampling framework is effective for signal reconstruction given a fixed energy budget [12, 11].

Adaptive GPS Trajectory Compression

In this project, I investigated the feasibility of compressive sensing for trajectory data reconstruction [2]. We case studied trajectory datasets from animal, pedestrian and vehicle and demonstrated that compressive sensing offers better signal recovery. The main contribution of this work was the dynamic adaptation of the compression subject to the speed of the object. We showed that compressibility is correlated to the speed of the object, however, compressive sensing theory does not offer any mechanism to adapt the compression. We introduced a cut-down version of support vector regression to enable an in-situ adaptation of the compression ratio.

Data Driven Construction of Projection Matrix for better Trajectory Reconstruction

Compressive sensing theory has been established around a predefined set of sparsifying (dictionary or or- thonormal basis functions) and measurement matrices. However, custom dictionaries and projection matrices can offer better sparsity offering better signal reconstruction accuracy. I applied the dictionary learning mechanism, which is known as sparse coding to obtain highly compressed representation of a signal. I also proposed a novel formation of measurement matrix by applying a special Singular Value Decomposition (SVD) on the dictionary. I showed that this combination substantially outperforms the predefined projection matrix and dictionary defined by the compressive sensing theory [8].

Gait Speed Prediction from room-to-room Transition Time

In this project, our objective was to extract human gait speed using non-invasive and cost-effective sensors. I worked with my collaborators in the Oregon Centre for Aging and Technology (ORCATECH), who are recognized worldwide for their development of technologies for the aging population. We have shown that [5] the room-to-room transition time preserves enough information to infer gait speed. This is an important finding, which can potentially save the cost of installing state-of-the-art sensor array that ORCATECH is currently using in their participant’s dwelling.

Physical Activity Classification

In this project, we attempted to infer physical activity, such as, walking, running, sitting and standing from smartphone accelerometer feed. In particular, I used l1-minimization, which is the key driver of compres- sive sensing, for classification and compared it with the classical support vector classification. I preferred l1-minimization, as it has many benefits over support vector classification. The most promising aspect of l1-classification is that it is feature-insensitive as oppose to support vector classification. Furthermore, l1- minimization does not need extensive training to build a model. Therefore, the classification process can be readily updated with the arrival of new class. These two attributes make the classification using l1 very attractive for resource improvised embedded platforms, such as smartphones, wireless sensor nodes etc.

Cost-effective Comfort Modeling in the Commercial HVAC buildings

In this project, I looked into human comfort modeling in HVAC controlled office buildings. Thermal comfort is very important for human health and productivity. However, to date most of the HVAC control strategies only focus on maximizing energy savings and underestimate comfort preservation. One of the reasons for not accounting for comfort is the need to measure complex personal health parameters, such as, metabolic rate, clothing insulation etc. I have shown that humidex (a combination of humidity and temperature) possess sufficient information to predict thermal comfort of individuals [7]. I have developed optimal HVAC control

strategies that control the temperature and humidity subject to preserving thermal comfort of occupants. In particular, I solved an optimization problem where the objective was to minimize the difference between indoor and outdoor temperature and humidity and the constraints were individual’s thermal comfort ranges. 

  1. [1]  C Chou, W Hu, et al. Efficient computation of robust average of compressive sensing data in wireless sensor networks in the presence of

    sensor faults. 2012.

  2. [2]  Chun Tung Chou, Rajib Rana, and Wen Hu. Energy efficient information collection in wireless sensor networks using adative compressive

    sensing. In Proceedings of LCN 2009: The 34th IEEE Conference on Local Computer Networks, pages 443–450, Zurich, Switzerland, 2009.

  3. [3]  Shahriar Nirjon, Robert F Dickerson, Qiang Li, Philip Asare, John A Stankovic, Dezhi Hong, Ben Zhang, Xiaofan Jiang, Guobin Shen, and Feng Zhao. Musicalheart: A hearty way of listening to music. In Proceedings of the 10th ACM Conference on Embedded Network Sensor Systems, pages 43–56. ACM, 2012.

  4. [4]  Cuong Pham and Tu Minh Phuong. Real-time fall detection and activity recognition using low-cost wearable sensors. In Computational Science and Its Applications–ICCSA 2013, pages 673–682. Springer, 2013.

  5. [5]  Rajib Rana, Daniel Austin, Peter G Jacobs, Mohanraj Karunanithi, and Jeffrey Kaye. Passive in-home estimation of gait velocity from motion sensors. arXiv preprint arXiv:1310.4880, 2013.

  6. [6]  Rajib Rana, Chun Tung Chou, and Wen Hu. Energy-aware sparse approximation technique (east) for rechargeable wireless sensor networks. In Proceedings of the seventh European Conference on Wireless Sensor Network (EWSN 2010), pages 306–321, Coimbra, Portugal, 2010.

  7. [7]  Rajib Rana, Brano Kusy, Raja Jurdak, Josh Wall, and Wen Hu. Feasibility analysis of using humidex as an indoor thermal comfort predictor. Energy and Buildings, 2013.

  8. [8]  Rajib Rana, Mingrui Yang, Tim Wark, Chun Tung Chou, and Wen Hu. A deterministic construction of projection matrix for adaptive trajectory compression. arXiv preprint arXiv:1307.6923, 2013.

  9. [9]  Rajib Kumar Rana, Chun Tung Chou, Salil S. Kanhere, Nirupama Bulusu, and Wen Hu. Ear-phone: an end-to-end participatory urban noise mapping system. In IPSN ’10: Proceedings of the 9th ACM/IEEE International Conference on Information Processing in Sensor Networks, pages 105–116, New York, NY, USA, 2010. ACM.

  10. [10]  Abdikarim Sheik-Mohamed and Johan P Velema. Where health care has no access: the nomadic populations of sub-saharan africa. Tropical medicine & international health, 4(10):695–707, 1999.

  11. [11]  Yiran Shen, Wen Hu, Rajib Rana, and Chun Tung Chou. Non-uniform compressive sensing in wireless sensor networks: Feasibility and application. In Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), 2011 Seventh International Conference on, pages 271–276. IEEE, 2011.

  12. [12]  Yiran Shen, Wen Hu, Rajib Rana, and Chun Tung Chou. Nonuniform compressive sensing for heterogeneous wireless sensor networks. IEEE SENSORS JOURNAL, 13(6):2120–2128, 2013.

  13. [13]  Qing Zhang, Mohan Karunanithi, Rajib Rana, and Jiajun Liu. Determination of activities of daily living of independent living older people using environmentally placed sensors. In Engineering in Medicine and Biology Society (EMBC), 2013 35th Annual International Conference of the IEEE, pages 7044–7047. IEEE, 2013.