Effect of Smartphone App on Activity

The aim of the study is to increase patients' physical activities by using a dedicated cellular application that will encourage patients to adhere to their doctor recommendation on a personal basis.

Primary outcome In diabetic patients: measuring an increase in daily physical activity In cancer patients: improvement of quality of life in correlation with the level of physical activity

Secondary outcomes In diabetic patients: improved glycemic control as assessed by sequential blood tests for HbA1c.

The patients will fill quality of life questionnaires (SF36) at recruitment and after 6 months. After 6 months the patients will also fill a questionnaire about their experience of using the app.

Each recruited patient will have an Android based smart phone. Each patient will provide:

1. Approval to join the experiment
2. Age, gender, height
3. Telephone number (for SMS)

Length of intervention - at least 6 months per patient. Each patient will be randomly assigned into one of two groups, which will specify feedback relative to himself or to others or a weekly reminder to exercise.

Number of patients:

1. Diabetes: 150 patients, of which 50 are controls.
2. Cancer: 100 patients, of which 20 are controls. All patients will receive instruction about the importance of physical activity and a personal recommendation for activity level, n sessions of activity per week, and time span per session (i.e., at least 2 hours of walking per week divided to 3 walking sessions per week) Patients in the treatment arms will receive at least n (number of commended sessions) messages per week of positive feedback if activity performed or negative feedback if not performed. At the chosen day each week the patient will receive a summary of the exercise for all the week.

Feedback Possible feedback

(NOTE - these the the actual feedback messages that the participants will receive, and are therefore in the second person):

1. Negative feedback: "You need to exercise to reach your activity goals. Please remember to exercise tomorrow".
2. Positive feedback:
3. Relative to self: "You're exercise level is higher than last week. Keep up the good work"
4. Relative to others: "You're exercising more than the average person. Keep up the good work"
5. Control arm: "Did you remember to exercise?"

Technical requirements

1. App - will collect physical activity and send it to a server. App will run in background without need to restart on reboot.
2. Server - Collects physical activity

Feedback policies The experiment will have two phases of feedback. Phase 1

The investigators begin with no data, so the policy at this stage is as follows:

1. Positive feedback will be sent each day if user has surpassed 1/7th of weekly activity that day.
2. Negative feedback will be sent every 3 days, if activity hasn't passed 1/7th of activity.

Each day, with a probability of 0.2, a random decision on feedback will be made.

This phase will last approximately 4 weeks. Phase 2 Using a learning algorithm (see below) the computer will adjust the feedback, and decide daily on the feedback (positive \ negative \ none).

Policy learning The investigators will start with a simple policy learning strategy, and later use more sophisticated methods that will have a state-space representation of the user.

The initial algorithm will represent each user at each day using the following attributes:

1. Demographics (age and gender)
2. Expected versus actual activity level this week (ratio of the two)
3. Last feedback given (positive \ negative)
4. Day of the week (we will use week-long cycles). The goal of the algorithm is to give feedback today so as to encourage activity tomorrow.

When training the algorithm, the computer will have a feature vector comprising of the attributes above, and a matrix of actions (for day t). The output to be predicted is whether the activity level on the following day (t+1).

There can be two types of feedback depending on weekly and daily behaviors:

Weekly goal Not achieved Achieved Daily goal (on day (t+1)) Not achieved 1 1+alpha Achieved 1+alpha 1 (alpha>0) The algorithm will pay a higher penalty if, for example, on a given day the message encouraged activity, but the weekly goal was not achieved compared to if it was.

For simplicity, the initial learning algorithm will be linear, until enough data is collected. That is, given a matrix:

X = (demographics, expected vs. actual activity, last feedback, day of the week, actions) And a vector showing the amount of activity on the following day, weighted as in the table above, denoted by Y, we will learn a vector of weights w such that: X * w = Y.

In phase 2 of the project the computer will use other learning algorithms. Exploration (random action at a given day) will continue throughout both phases at the same level.