redFlag

• 

RED FLAG Development of a multiparameter prediction algorithm for mania in persons with bipolar disorder.

Keywords: mental health, preventative medicine, behaviour analytics

Mahesh Jayaram (neuropsychiatry, research) Jahan Penny-Dimri (MBBS candidate, information technology) Xavier Taylor (computer science) Luke Perry (MBBS candidate, research)

Document Contents Summary of System Medical/healthcare/societal context in greater detail (if we have time) Technical Design of System Privacy, legal, ethical considerations Financial Competitors Accompanying research trials Potential Extensions

Introduction to Red Flag System RED FLAG is an early detection system for mania in persons with bipolar disorder that incorporates using geospatial, financial, social media, and telecommunications data as deviations from an individual's norm. We will investigate the optimal parameters and weightings in the predictive algorithm in a discovery phase (phase I) followed by investigating the clinical benefit of using the algorithm in a randomised-controlled trial (phase II). We will extend the scope of this software to predict relapse in other mental illnesses after proof of concept, acceptability, and proof of clinical utility of RED FLAG have been established. Medical/healthcare/societal context in greater detail (if we have time) The economic burden of bipolar disorder in Australia is over 5 billion dollars. Bipolar affects over 230,000 individuals. The cost of an acute inpatient admission during a manic episode is over $37,000, whereas the cost of a visit to the outpatient clinic at a crucial time is $80-250. RED FLAG will reduce the need for acute inpatient admissions and improve patient wellbeing by providing an early detection system that enables mental healthcare providers to contact their patients about their well-being before they even reach full mania. Technical Design of System System overview The Red Flag system uses, collects, and analyses data about users’ behaviours, calculating a real time probability that a patient experiencing a manic episode based on changes in this data. When the probabilistic threshold is passed, the system sends an alert email to nominated clinicians and carers, informing them that the patient may be in a manic episode. Most of the data is passively captured from the clients smart phone, with the exception of the patient's bank account information, which is directly accessed by the Red Flag server. The mobile data is sent to Red Flags server, where it is stored and analysed. The Red Flag app is installed on the patient’s smartphone, but runs in the background - there is no patient front end.

An integral feature of the system is the modularity of its constituent data sources and their analysis. While the accuracy of the probability estimate is greater if all possible data sources are included in the analysis, the system is designed to still deliver a useful score if only a subset of the data sources are available. This will allow flexibility with respect to variation in smartphone hardware, and also variation in the extent of consent provided by the patient. Mobile Application We intend to develop the Red Flag mobile application using the Appcelerator Titanium SDK. However, before proceeding with implementation we will more thoroughly investigate competitors to assess the most suitable framework. The application is installed on the patient’s phone, and sends data to the Red Flag server, requiring no actions from the patient. Specification of data collection functions:

Table 1 paints a big picture of the various functionalities that would together comprise Red Flag. The various data streams would be analysed in a variety of ways, depending on the nature of the data. The various data analysis outputs will be combined to produce a probabilistic estimate of whether the patient is experiencing a manic episode.

Table 1: Passive Predictive data capture categories and associated technologies Data Category Behavioural Change Technology for Data collection Geolocation Average distance travelled Locations visited Time of movement

For GPS based geolocation analysis, Titanium Geolocation module.

http://docs.appcelerator.com/platform/latest/#!/api/Titanium.Geolocation

Social Media Post frequency/length/time Grandiosity/Inflated self esteem Spelling Errors Message frequency, length, speed (in IM)

For sentiment analysis, we will use IBM’s alchemyAPI. http://www.alchemyapi.com/products/alchemylanguage/sentiment-analysis Telecommunications Number, length of outgoing communications Who is contacted Time of communications

Logging call and sms metadata. Smart phone (in)activity Decreased need for sleep Distractibility - increased overall use of phone Average session length in applications

Logging user initiated events, application initiation, switching etc Financial Magnitude, frequency of expenditure Time of transactions

Yodlee API for bank balance and transaction information.

https://www.yodlee.com/products/apis/aggregation-api/ Keystroke analysis Speed Autocorrect use

Logging keystroke events metadata across applications.

Data Analysis The specifics of the data analysis algorithms to be selected for the various data sources have not yet been decided. In some cases, a fairly simple calculation, such as mean expenditure per week or month can be calculated, and deviations of a given percentage given a particular weighting. Other data might be analysed in a more complex manner, for example movement patterns might be clustered or classified by various unsupervised, or (respectively) supervised machine learning algorithms. The truth is that this project is fairly complex, and the need for further research of various specifics, especially with respect to medical statistics, probability etc, is essential before proceeding with implementation.

The overarching principle is of detecting statistically significant variations (usually in one direction, but often in either), which indicate the presence of one or more symptoms of mania. When the probabilistic threshold is exceeded, the nominated professionals and carers(if any) are notified via email.

Privacy, legal, ethical considerations

The acquisition of data on a patient's travel, financial, and communication habits raises significant privacy, legal and ethical considerations. We will protect users’ privacy by ensuring the data collected from their smartphone is encrypted, de-identified, and stored in a secure virtual space. There will be no permission or access to this data by any party without express written informed consent by the user. Specifically, data on telecommunications and financial expenditures will not be accessed by any party. The user may specify which type of data is collected, and which data is shared with their healthcare worker, which may range from alert only, to enabling access to a broader dataset. At any stage, users may opt out of RED FLAG and request that their data be destroyed at no penalty.

We will consult relevant legal professionals to best establish a framework that protects absolute privacy and patient autonomy.

Financial Because of its modular nature, Red Flag is well suited to agile development, with independent functionalities being shippable in a piecemeal manner.

We will seek funding from governmental research bodies, charitable bodies (eg Stanley Foundation), and through crowdsourcing. The operational costs will cover software development, server space, and research costs (eg research nurse for longitudinal discovery phase (phase I)). Preliminary estimates of operational costs are $5,000-10,000 for the longitudinal development phase. The cost of rolling out the final product will be informed by the results of phase I.

An industry professional with experience in projects of this scope estimated that the full development of a software system of this scale might cost within $200,000 and $500,000. Competitors We have identified other published systems that attempt to predict mania in persons with bipolar. Each system only incorporates parameters from a limited number of domains (eg keystroke analysis), and does not attempt to aggregate geolocation or financial data (novel in RED FLAG). Many also require active patient participation, for metrics such as self reported mood.

BiAffect: an android app that uses keystroke dynamics to track mood in bipolar disorder and to detect mania. Currently still in development phase, but was a finalist in the moodchallenge hackathon. http://www.moodchallenge.com/biaffect-finalist/

A systematic review of mobile applications for bipolar disorder included 32 applications and found that none had undergone extensive evaluation through research (Nicholas et al, 2015). A consistently deficient feature was a lack of privacy policies; a fact underscoring the importance of effective privacy planning with RED FLAG. None of the included apps incorporated the breadth of input parameters that RED FLAG does.

Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4642376/

By requiring no active input from patients, Red Flag will avoid entirely the perennial issue of low patient compliance to reporting and treatment regimes.

Accompanying research trials We will investigate the optimal parameters and weightings in the predictive algorithm in a discovery phase longitudinal study (phase I) followed by investigating the clinical benefit and acceptibility of RED FLAG in a larger population in a randomised-controlled trial (phase II).

Potential Extensions Once the acceptability and utility of RED FLAG is established in bipolar disorder we will endeavour to extend the scope of this system to predict relapse in other mental illnesses such as schizophrenia as well as to predict suicidal behaviour and suicide completion in a diverse population people living with mental illness.