Chapter 2 Related Work
Mental Health Assessment and Intervention
Monitoring mental health has been a core component of modern clinical settings. The Diagnostic and Statistical Manual of Mental Disorders (DSM) is an evolving diagnostic tool that has become the mental healthcare professional’s standard for psychiatric diagnosis in the USA . The Clinical Global Impressions-Severity (CGI-S) scale was developed by mental healthcare professionals as a simple measure for the severity of a patient’s symptoms , and has been used as a measurement tool for cases of depression in the past . It has since been evaluated for its value in clinical settings with positive results . It is a simple scale from 1-7, where higher ratings signify more severe symptoms and negative effects on the patient’s health. Forming the initial response to someone’s crisis or mental illness is also critically important. Today’s counselors and psychotherapists continually develop the best response and treatment of a person at risk of suicide . The key role that the general public can play in this scenarioisencouragepeopleathighrisktowardsqualifiedmentalhealthsupportservices. In communities of young students, suicide prevention training programs such as “Ask, Listen, Refer”havebeenimplementedtoincreasepublicawarenessofriskfactorsandtogiveinstruction on how to refer people to professional care . Another variant, “Question, Persuade, Refer” (QPR), was founded for the purpose of empowering trainees to be active, rather than passive, when encouraging someone else towards support [22, 31]. While no study today has been able to directly measure the number of prevented suicides through QPR training, approximately 90% of completed suicides are people lacking treatment for mental health disorders (including depression) , and it is widely agreed that proper treatment would save lives .
Expressive Writing Therapy
Emotional confession and disclosure have been an interest for understanding the psychological healing process . This has been standardized into expressive writing therapy (EWT), a writing task that has been studied in a multitude of settings and populations with results that vary . EWT’s resemblance to a personal journal is striking due to the expressive freedom and introspective writing focus. Groups with depressive symptoms who engage in EWT report decreased depressive symptoms, and increased life satisfaction . While EWTdoesnotnecessarilyreduceintrusivethoughts, itcanmoderatetheimpactofintrusive thoughts on depressive symptoms . The reduction of depressive symptoms has also been reported in college students after a time delay .
Natural Language Processing
Natural language processing (NLP) is the transformation of normal human language into measuresthatcanbedirectlyusedbycomputers. Ith as commonly been used to translate one humanlanguagetoanother,ortointerpretauser’squeryinavalidformatforacomputation. The linguistic characteristics a person uses (verbal or written) can even give insight on their psychological state  and on their feelings towards a topic . “Linguistic Inquiry and Word Count” (LIWC) is an example of a NLP tool purposed for quantifying text along 93 different linguistic and psychological measures, such as “positive emotion” and “certainty” [30, 36]. For college students writing essays, some of the LIWC measures correlated to whether they were currently depressed, formerly depressed, or had never been depressed .
Machinelearning(ML)isaproblemsolvingmethodofcomputerstocompleteacomplextask without relying on specific instructions. This typically requires input data that has some relationship (however complex) to the solution of a problem. The relationship between the input data and the solution can be computed or inferred through many different approaches (The models we will use are detailed in Section 4.4.1). Useful data for understanding a person’s health is widely available on a person’s computer, and collecting this data can be done without interfering with the device’s normal use . Machine learning has been utilized on the data of a person’s smartwatch to identify healthrelated behaviors with 93% accuracy, an improvement over smartphone data yielding 77% accuracy . This serves as some evidence that the more connected a source of data is to the target (the problem’s solution), the more accurate predictions will be given by the machine learning model. Another study produced machine learning models that use a person’s Twitter usage and post content to predict the future onset of depression for those users with 70% accuracy . These examples show the usefulness (and some limitations) of using machine learning for the purposes of understanding and improving mental healthcare.
Chapter 3 Proposed Solution
It is beneficial to establish an ideal solution as a reference when making the design decisions for ALJI. The enables us to prioritize the key aspects over minor concerns and anticipate conflicts that would diminish our possible impact. ALJI is meant to be an alternative (or addition) to a personal journal. At all times, the author needs to remain aware that ALJI is not a replacement or an alternative to a counselor or psychologist.
ALJI must be created in a way that doesn’t sacrifice the benefits that EWT provides (see Section 2.2). For this reason, we should match the natural journaling task and environment of an author as closely as possible. Journal authors expect complete controloverthe privacy and accessibility of their journal. ALJI needs to win an author’s trust by being completely transparent and under the control of the author. The best way to ensure this is to create ALJI as a standalone, normal computer program without the capacity to communicate with any other device. At any time, authors can erase all traces of ALJI and their writings in the same way they can burn a journal. An online service would allow journal entry communications to be intercepted, and the service itself can be a target for data breaches that are commonplace on the internet. Relying on an online service would not be acceptable for journals that need to never be copied, accessed, or shared without the author’s permission. Authors also need to see the entire inner workings of the ALJI program and verify our claims of privacy, which leads us to make ALJI as open source software and prefer open source dependencies. The entire history and current code for running ALJI is publicly accessible at this GitHub repository: https: //github.com/sublime09/ALJI. If ALJI were to fail in these requirements, then authors would understandably lose trust in ALJI. A breach of trust fundamentally changes how the author would write journal entries, viewing them more as social media posts or an interview environment where their character is being openly judged. This change in environment can remove many benefits of EWT and negatively impact the power and accuracy of ALJI’s predictions of the author’s mental wellness.
Figure 3.1 shows the basic interaction sequence of the ALJI. ALJI is composed of three modules that interact with the author (or user): the Machine Learning Module, the Crisis Criteria Module, and the Referral Module.
Machine Learning Module
This module in fig. 3.1 contains the machine learning model that is best suited for quantifying and predicting the author’s mental health. This model will accept the text of a journal entry, standardize the text (e.g. segmenting, parts-of-speech tagging, stemming) to an acceptable form, analyze the text for basic psychological markers, and then output mental health predictions. Note that the machine learning model is already trained at this step. The kinds of predictions that are made depends on what the model is trained to predict: tiredness, weight changes, substance abuse, insomnia, catastrophic thinking, thoughts of death, self-harm, suicidalideation, suicideplanning, oranymetricthatisdeemedimportant towards understanding the mental health of someone. Uncertainty around a mental health indicatorneedsfurtherinvestigation, andsothatmentalhealthindicatorshouldbeincluded in the predictions. However, these predictions represent the possibility of mental health indicators, and should not be taken as truth.
Crisis Criteria Module
This module from Figure 3.1 assists in confirming the accuracy (or inaccuracy) of the Machine Learning Module. Meeting the “crisis criteria” of this module is done though a single question: if the previous predictions were taken as truth, would this justify an intervention on behalf of the author? It is vitally important that mental health professionals determine this criteria, not software designers or the general public. If the criteria are not matched, then ALJI has no further work, since all mental health indicators possibly detected did not warrant an intervention. Meeting the crisis criteria means ALJI must reduce the uncertainty around the author’s mental health before proceeding. Human-computer miscommunication is incredibly common,especially when the text input from journals allows for the author’s expressive freedom. ALJI gives direct yes/no questions to the author, asking if they are currently experiencing the mental health indicators that were predicted. It may be possible to quote meaningful passages in the author’s writing sasareference for whyALJIisaskingthesehealthquestions. These clarification questions are a defense against misunderstandings due to sarcasm, humor, hyperbole, non-literal language, and fictional events in the journal entry. The author’s response to these yes/no questions is taken as absolute truth, and compared against the same crisis criteria. If the criteria are matched again, then the Referral Module is triggered to intervene and guide the author towards mental health support services. No software is perfect, but in the context of ALJI, we prefer certain errors over others. Consider a false positive prediction in coming from the Machine Learning Module: the author is healthy ,but the prediction raises concerns. This can easily be corrected by the clarifications in the Crisis Criteria Module. Even if it was not, the author is given a referral to mental health professionals that are capable of understanding the error. A much worse scenario happens when a false negative error in prediction occurs: an author that needs guidance towards support, but it is not given to them. Thankfully, many models are able to leverage a trade-off between a critical error and a non-critical error by assigning weights to certain samples [3, 4]. A crude version of this could be achieved by simply duplicating samples that we wish to have more weight within the Machine Learning Module.
This module in Figure 3.1 provides the author with the intervention and guidance towards appropriate mental health support services. It would be most persuasive if this guidance presented reasoning or new information to the author about the concern. Which support services it shows depends on the mental health indicators that are confirmed by the author. Forexample,indicatorsforconsistentdepressedmoodandrecentlossofinterestcouldresult in a response of what depression is, how treatable it is, and how to make an appointment at a counseling service nearby. Some mental health concerns require an immediate response, so every step of ALJI must be swift. An appropriate response to suicide planning and the tools for suicide would be providing the contact information of a 24-hour national suicide prevention hotline and the medical emergency telephone number. It is tempting to design ALJI so the most serious mental health concerns are immediately and automatically report to emergency services, with no input from the author. However, this would break the author’s trust of privacy and nonjudgement, leading to the same consequences outlined in Section 3.1 and lessening those positive impacts of ALJI. It also ignores how receptive the author would be of outside aid at that moment, or how they may need a moment to decide for themselves. Forcing aid upon someone who does not want it can have absolutely dire consequences. Designers of ALJI should always be aware that ALJI is an alternative to pen-and-paper journals, and that journal authors could easily switch to a writing medium that respects privacy. The more that ALJI is removed from the environment of a private journal, the fewer journal authors will consider using ALJI.
List of Figures
List of Tables
2 Related Work
2.1 Mental Health Assessment and Intervention
2.2 Expressive Writing Therapy
2.3 Natural Language Processing
2.4 Machine Learning
3 Proposed Solution
3.1 Human-Centered Design
3.2 Interaction Sequence
4.1 Gathering Journal Data Source
4.2 Measuring Journal Language
4.3 Journal Labeling Participation
4.4 Machine Learning Process
5.1 Basic Journal Results
5.2 Journal Labeling Results
5.3 Learning Models Results
6.1 Journal Data Source
6.2 NLP and Empath
6.3 Journal Labeling Process
6.4 Learning Model Outcomes
7 Future Work
7.1 Expanded ALJI solution
7.2 Journal Source Improvements
7.3 Enhanced NLP
7.4 Journal Labeling Expansions
7.5 Machine Learning Modifications
7.6 Clinical Study of ALJI Prototype
7.7 Clinical Variation
GET THE COMPLETE PROJECT
ALJI: Active Listening Journal Interaction