Research Topics and Example Applications of Smart Healthcare
Wireless sensor networks are applicable in many different healthcare settings, each presenting its own unique challenges. We are pursuing a variety of interesting research topics: some that underlie the entire field and some that are specific to an application. As with all systems research, some issues are not exposed until theory turns into practice, so new topics will be listed as they are explored.
Cardiac Health
Cardiac arrhythmia is any change from the normal beating of the heart. Abnormal heart rhythms can cause the heart to be less efficient, and can cause symptoms such as dizziness, fainting, or fatigue. Since they are sometimes very brief, it can be difficult to properly characterize them. Cardiac stress tests attempt to induce the event while the patient is wearing sensors in a laboratory.
In a homecare setting, wearable EKG sensors can monitor for the condition continuously, over days or weeks, until the event occurs. The recorded data is promptly sent to the physician for analysis. If the event is serious enough, the emergency communication channel may be used to call for help, or it may be dispatched automatically. Other sensors may be able to record environmental data to help identify aggrevating factors (side-effect of medication, little sleep, etc.).
Circadian Activity Rhythms
CARs reveal intimate details about a person's living activities and health status. As WSNs grow in their capability to collect, process, and store data, personal information privacy becomes a rising concern. Our system includes a framework to protect privacy and still support timely assistance to patients with critical health conditions.
Data access rules can be dynamically altered based on contexts generated by the CAR algorithm when necessary. For example, if a patient has blocked access to their ECG data for nurses, but the CAR has determined serious anomalous behavior that might indicate a heart problem, then the nurse is alerted and access to the data is allowed for a period of time. In case of an alarming health status, alarms are sent over the network to restrict or relax the privacy depending on the user's role. Other contexts in our system include the patient's physiological conditions (ECG and pulse readings), living environment conditions (room temperature and light), and autonomy (inferred from ADLs by the CAR).
Context-Aware Power Management
Circadian Activity analysis provides behavior patterns of the resident, from collected sensor data and predictions. Power management in AlarmNet adapts to the behavior of the resident. Users can directly control sensing functions of sensors in the network using an interface on a PC or PDA. These commands are sent to the sensor nodes via the hierarchical routing protocol. Users can also define context policies. From these policies, power management commands are automatically generated by the AlarmGate software on the backbone and forwarded into the network.
Journaling Support
Journaling is a technique recommended for patients to help their physicians diagnose ailments like rheumatic diseases. Patients record changes in body functions (range of motion, pain, fatigue, sleep, headache, irritability, etc), and attempt to correlate them with environmental, behavioral, or pharmaceutical changes.
The system can aid patients by:
- providing a time-synchronized channel for recording and transmitting the journal (PC, PDA, dizziness button),
- recording environmental data or external stimuli (temperature, barometric pressure, sunlight exposure, medication schedule), and
- quantitatively measuring changes in symptoms (pain, heart-rate, sleep disruption).
Sleep Apnea
Every night the system monitors blood oxygenation, breathing, heart rate, EEG, and EOG using on-body sensors to assess severity and pattern of obstructive sleep apnea. Bed sensors monitor agitation (movement) and store and report sensor data. The provider and patient are alerted if oxygenation falls below a threshold. Monitoring can continue while treatment efficacy is being assessed.
Multi-Modal Data Association
Data association is a way to know "who is doing what?" in a system with multiple actors present, such as an assisted-living community. It permits us to recognize the right person when he or she is responsible for a triggered event. This is indispensable if we want to avoid medical errors in the future and attribute the appropriate diagnostics to the right person.
For endo-sensors (sensors worn on the body), data association is made easier since an ID can be bound to the sensors and sent over the network within the data packet. This solution can be easily implemented with any body networks and may also be applied to any exo-sensors (motion, etc.) using an
However, a problem still remains: what happens when the body network/transmitter with active ID is not worn? Is the person still recognizable? Integrating an active ID in the environment is quite challenging. Here, we have to invent new solutions to address that more complex problem. One approach is to associate different percentages of "recognition confidence" to different kinds of residents (patient, caregivers, family, friends, doctor, etc.). These estimations should be sufficient for long-term data-analysis, where some noise from misclassifications may be tolerated. But for short-term monitoring, it is not sufficient. For example, vital signs must be attributed to the right person, especially if they generate emergency alerts.
The infrastructure of the network must embody a set of rules about how to distinguish among multiple people present in a single space. It must know both the number and identity of persons present, using techniques such as: video recognition, speech recognition, "magic carpet" for gait recognition, Circadian Activity Rhythms (for medium pattern recognition), elementary pattern constraints among low-level sensors (e.g., two consecutive weights cannot change a lot), and fusion of all of these techniques.
Data Integrity and Cleaning
When the data association mechanisms are not sufficient, or integrity is considered critically important, some functionalities of the system can be disabled. This preserves only the data which can claim a high degree of confidence. In an environment where false alarms cannot be tolerated, there is a tradeoff between accuracy and availability.
Impossible scenarios or anomalies discovered using data redundancy can also be removed from the database, or be flagged as problematic. Thus, the data is proactively "cleaned", rather than allowing the gradual accumulation of fault-producing data.
Security and Privacy
The system is monitoring and collecting patient data that is subject to privacy policies. For example, the patient may decide not to reveal the monitored data of certain sensors, for example, the pulse, until it is vital to determine diagnosis and therefore authorized by the patient at the time of visit to a doctor.
Such “delayed revealing” requires storage of the data without access to it until authorization is granted. Alternatively, patient may decide not to reveal his data as an individual, but may agree to participate in a public study of a number of individuals. Such data is tagged as "anonymous" and is de-identified for study purposes.
Emergency-aware applications demand a privacy protection framework capable of responding adaptively to each patient's health conditions and privacy requirements in real-time. Therefore, traditional role-based access control which makes access authorization based on users' static roles and policies is not flexible enough to meet this demand. We implement a privacy protection framework which is dynamically adjustable to users' context, allows data access authorization to be evaluated on the fly, and is able to adapt to patients' emergency cases.
Security mechanisms on constrained sensor devices must be lightweight with respect to their storage, computation, and communication costs. We have developed and are refining a hardware-accelerated secure communication layer for the MicaZ motes. It provides support for multiple keys among neighbors and application-selectable protections.
Real-Time Data Streaming
Since sensor networks deal with real world situations which often require a timely response (e.g., medical emergencies, vital sign reporting, tracking), it is necessary to provide real-time data streams. Due to resource limitations of both computation and communication, and unpredictability in network topology and work load, it is impractical to guarantee hard real-time constraints. We develop our solutions based on experience and knowledge of these constraints to provide probabilistic guarantees for the system's timing requirements.