Design Concept

This report explores the concept of a new application called The ‘Travelling Clinic’ that uses built-in mobile phone accelerometers to record patients‘ movements. TC uses the mobile phone's memory storage and the ability to transmit data wirelessly to share their reports with their doctors/nurses and even trusted family members on a regular basis (the reporting time period can be set by patients). The application has a simple interface that allows patients to view easy to comprehend visualizations of their mobility records in different weeks and months of the year.

It speed dials a trusted family member in cases when it senses a fall. It automatically shares mobility records, visualizations and progress reports with the patient's doctor/nurse and trusted family member. Thus, it helps keep track of the disease and its
stages. Such close tracking can help provide the patient with timely medication/help and prevent the condition from worsening. Travelling Clinic aims to help patients lead independent lives without actually having to travel to clinics frequently. In cases when the patient might be living alone, the application keeps a trusted family member informed about the patient's mobility records and response/progress to medication. Below explained are all the interface stage diagrams followed by their explanations. I used the tool ‘Sketch’ to design the interfaces.

Working of 'Traveling Clinic' - Background and Generating the Training Data Set (Support Vector Machine)

TC uses machine learning techniques to track movements of patients carryingmobile phones. TC is inspired from the research carried out by Mark V. Albert,Santiago Toledo, Mark Shapiro and Konrad Kording on the topic ‘Using mobilephones for activity recognition in Parkinson's patients’. They first built atraining set for their support vector machine algorithm. They achieved this by8recruiting eighteen healthy individuals and eight patients suffering from PD. Alltheir subjects were instructed to carry T-Mobile G1 phones running AndroidOS version 1.6 in their front pockets. These phones have a standard built-intri-axial accelerometer with a range of plus-minus 2.8 g. The sampling ratewas variable between 15 and 25 Hz depending upon the amount of movement.Subjects were instructed to perform a number of different activities, each for at least 1 min. Before each activity, the subject would select the activity on aspecially designed phone app with the experimenter present to minimize errors. 

The accelerations were labeled according to the activity they were performing.They took into account different phone orientations as well by generating threeadditional samples for each recorded sample using a coordinate transform thateffectively flips the phone 180 along each of its three axes. TC can use thesame technique to generate its training dataset.Another interesting addition that TC has made is its ability to detect patientfalls. In order to generate an additional training data set for this, I suggesttaking a mannequin and putting the phone in his pocket. After this, we need toartificially simulate different possible fall velocities and positions by allowing themannequin to fall in different ways. The acceleration readings are noted downin every case in order to generate the training data set for falls. The idea ofgenerating the ‘fall’ training dataset is inspired from Wibisono's work. Theysuccessfully used a threshold based fall detection algorithm that processes datafrom triaxial accelerometer and magnetometer in phones in order to detect falls.The algorithm used Signal Vector Magnitude (SVM) peak value, base lengthand post-impact velocity to distinguish falls from most of daily activities.However, the SVM curve in a period produced by running is similar to a fall.Thus, the vertical acceleration was also observed by them in order to increasedetection accuracy and differentiate the two motions. In their experiments, thedata was collected by simulating fall in four directions: forward, backward, leftand right.

Recruitment of patients for generating records

Parkinson patients can be recruited for clinical research using this platform:CenterWatch ( https://www.centerwatch.com/clinical-trials/listings/condition/117/parkinsons-disease/). Wake Research Associates (Raleigh,NC USA) engage in clinical studies involving Parkinson’s patients as well. Wecan get in touch with them and possibly collaborate.

SVM Applied to 'Traveling Clinic'

We use the training datasets (both for regular activities and for detecting falls) that we generated through the study discussedabove. We first begin by finding the hyperparameters using a grid search. Thehyperparameters for SVM are the kernel function which depends on a soft marginconstant C and other parameters e.g. width of a Gaussian kernel or degreeof a polynomial kernel. For a large value of C, a large penalty is assigned toerrors/margin errors. When C is decreased those points become margin errors;the hyperplane's orientation is changed, providing a much larger margin forthe rest of the data. The grid search method is a popular method used inmachine learning for optimization of hyperparameter values.The hyper parameters are found using a grid search of 10x where x is an integerbetween 5 and 5 and selecting the maximum cross-validated error in predictingthe healthy subject labeled activities. For SVM, we normalize each featureto have 0 mean and unit variance. We then apply radial basis functions, givingus two hyper parameters the soft slack variable, C, and the size of the Gaussiankernel, Lambda. The values found by cross validation should be around C =1 and Lambda = 0.1 for the across subjects'validation and C = 10 and Lambda= 1 for the 10-fold validation (discussed below).

10 Fold Cross Validation Method 

First, the goal is to compute a classification accuracy measure that can be comparedacross studies. To do this, a 10-fold cross validation, selecting every 10thsample for the test set. This accuracy is expected to be fairly high consideringmovement patterns specific to individual subjects were in both the training andtest sets. For SVM classification, this lead to a 96.1 percent accuracy forhealthy subjects, and a 92.2 percent accuracy for PD patients in Konrad’s study.

Reading Sensor Values, Data Visualizations and MobileApplication Development (Work Flow)

We can use an app development platform such as Appery.io to build an androidand an ios mobile application as discussed above. The mobile accelerometer readings canbe read using MATLAB. One can simply use the following command in Matlab‘sensor = sensorgroup(‘AndroidMobile’)’ after establishing a connection withthe phone's accelerometer. One can also use the function ‘sensor.Orientation’to get accurate readings. These readings are then classified into one of thesethree categories - ‘Walking’, ‘Standing’ and ‘Fall’ using the SVM algorithm (asdiscussed above). This categorized data is then stored on a cloud platform.

One can then use D3.js to build daily, weekly and monthly visualizations fromthis data. These visualizations are displayed on the interface (as shown above).We only use two types of simple visualizations - bar graphs and line charts.In case a fall is detected the system should flag the record and auto-call aclose contact. The family member's contact information as well as the doctor'scontact information (taken in as inputs from the settings page) should be storedin a separate database (link it to the records database) on cloud.