Title:A SYSTEM TO ASSIST CLINICIANS MONITOR THE CLINICAL STATE OF PATIENTS WITH PSYCHOSIS
Psychiatric patients, such as those suffering from depression or schizophrenia, often need to be monitored with frequent clinical interviews by trained professionals to avoid costly emergency care and unfortunate events (e.g., suicide attempts). However, there are not enough clinicians to monitor these patients on a regular basis. Furthermore, intermittent clinical evaluations are less accurate than regularly scheduled interviews and clinicians may miss subtle changes in patient state that occur over time. These limitations can affect both the quality and the timeliness of treatment. The primary goal of this project was to incorporate novel technologies to track clinical states of psychiatric outpatients longitudinally and, when appropriate, alert clinical staff to contact the patient. The research focused on three questions: (i) Can mobile technology be used to collect regular performance measures from patients? (ii) Can unobtrusive interactions such as speech and gestures from patients be converted to measures of patient state? (iii) Can this technology be configured to monitor a patient's clinical state with sufficient accuracy as to be useful?
The project developed a number of innovations to address the research questions:
-An iOS application that allowed patients remotely to easily self-administer daily interactions through a smartphone.
-The application engaged patients in spoken and touch-based interactions to capture structured daily measures of cognition, motor skill, and language.
-The recorded speech samples were automatically analyzed using recent technological advances in speech and language processing to recognize the content in the language and patterns in the speech.
-Multiple measures extracted from each patient's interaction were combined using machine learning-based techniques to accurately model clinician judgments and derive sensitive and reliable indicators of psychiatric patients' clinical state.
-Measures were analyzed longitudinally in order to track subtle changes in patient state that could be used to alert clinicians.
To develop the system, the research team first conducted a user needs study, in which they surveyed clinicians to assess current assessment practices and gaps. This study examined which types of clinicians conduct assessments, how frequently assessments were administered, limitations of the assessment approaches, and key types of measures clinicians would use in order to help with decision-making about patient states. Based on the study, the project team devised 25 unique behavioral assessment item types that assessed cognition, motor skill, and language. The items were similar in form and structure to standardly employed neuropsychological tests, but were designed so that 1) the items could be used for daily and remote administration with a smartphone, 2) the items provided short engaging tasks that required the patients to listen, watch, speak, and touch to interact with the smartphone, and 3) the interactions could be automatically analyzed to extract a rich set of measures from each item. Usability tests were conducted on the application to ensure that patients could remotely self-administer the test without the need for supervision from clinicians. A server-based scoring system was developed to analyze the speech and actions and generate predictions of changes in patient state.
Three data collection trials were conducted with 134 patients with schizophrenia, schizoaffective disorder and bipolar disorder, and 219 demographically comparable healthy participants. Two of the trials occurred in the United States and one in Norway. The trials served to iteratively refine the application, improve the accuracy and reliability of the assessment items, and validate the performance of the system.
Results demonstrated that the system was easily usable by both patients and the non-clinical participants. Speech processing and machine-learning based models built on the data showed that these models could accurately characterize important features associated with clinician judgment of patient states over a range of items. For example, the system could predict participant emotional valence and arousal as accurately as human raters. The system could further measure patient verbal recall of stories and generate accurate measures of patient memory and changes over time. Along with providing reliable measures, the techniques developed provide deeper insight into the role of cognition and language in clinical disorders. For example, patients showed less coherence in verbal recall as measured by automated techniques. Overall this project resulted in a novel assessment approach and a system incorporating state-of-the-art technologies for automatically monitoring patients remotely which can be used to track subtle changes in clinical states and alert clinicians.
Current best practice in assessing psychiatric clinical states relies on frequent clinical interviews with trained professionals. However, there are not enough clinicians to monitor patients as often as necessary, and relapse or suicide can follow even a few days after a clinician-patient meeting, highlighting the limits of intermittent clinical evaluation. Emerging technologies can automatically track and transmit key patient data, including information about affect, activation level, and suicidal ideati on, and generate alerts to initiate human intervention. Data capture can be inobtrusive (e.g., wristband monitors and smartphones). Body movement and non-verbal aspects of speech can form robust indicators of psychomotor activation and affect, while autom atic content analysis of speech can detect morbid ideation and changes in symptoms of depression and schizophrenia. We propose to assemble a remote monitoring system from currently available state-of-the-art technical elements, configure them to capture a nd transmit both passive-continuous and elicited-episodic data streams from outpatients. The goal is to implement a system that tracks the psychological state of psychiatric outpatients and when appropriate generates alerts that indicate that clinical sta ff should initiate communication with that patient. The potential to predict relapse risk before it occurs can avoid many medical crises and suicide-attempts/suicides. The research program is designed to answer three questions about the feasibility of rem ote monitoring of such patients: (i) Can available technology be configured to support such monitoring with sufficient accuracy to be useful? (ii) Will some portion of the outpatient psychiatric population tolerate such monitoring and comply with its beha vioral requirements? (ii) Can such monitoring be implemented effectively that better serves the outpatient population, while reducing overall costs?