USING MACHINE LEARNING TO INCREASE TREATMENT EFFICACY IN MENTAL HEALTH.
Something unique to every machine learning company is the precise nature of their hyperparameter optimization and goals of their model. We will optimize aifred with the help of a distributed network of domain experts in psychiatry -- a collaboration unique to aifred health. We are implementing attention networks responsible for removing the “black-box” nature of neural networks. As well, we are analyzing the quality of model predictions, allowing both for greater interpretability of model decisions and the generation of new basic research questions, which are going to be unique to the data-set and optimization techniques we develop in-house. By training aifred on reliable datasets, we are able to ensure quality input to our model. De-identified patient outcomes will feed back into our neural networks to continuously improve aifred’s predictive power. Feature engineering is an important part of determining which inputs go into a network and varies how it’s done for every team- once again, this will be undertaken with the support of diverse group of experts we are recruiting.
The aifred solution makes use of innovative and powerful machine learning techniques predict treatment efficacy based on an array of patient characteristics.
Forget the blackbox! Our system will provide a report highlighting the most significant features that led to a treatment prediction.
Patient Data Tracking
Track patient symptoms and test results to monitor outcomes or make new predictions. Banks of standardized questionnaires, data visualization, scheduling software -- all of it modular and capable of being tailored to clinicians' needs.
Electronic Patient Record
Keep all important patient information in one place, and get insights using our analytics.
FROM FUNDAMENTAL TO CLINICAL
Our research team is conducting a series of systematic literature reviews to curate predictors of treatment response and side effect burden in depression. We are evaluating the state of precision psychiatry in domains including genetics, endocrinology, immunology, metabolic biochemistry, and neuroimaging, as well as examining the feasibility of including biomarker testing in routine clinical practice. The results of these reviews will serve to validate our model and inform the input feature space by integrating these multimodal biomarkers along with sociodemographic and clinical factors.
Clinical research is focused on validating our model in controlled and real-world conditions. We are designing three kinds of research trials indicated below. Safety is critical, so our clinical team, which includes two physicians, will be making sure to review our model’s predictions and ensure hard-coded safety features so that model treatment recommendations are safe. We are blazing the trail when it comes to clinical validation of deep-learning based clinical decision aids, and as such are investing heavily in the development of ethical principles to guide development and testing. In fact, ethical development is so important to us that we have created our own ethical framework, known as Meticulous Transparency, to guide our work. We also never store personally identifiable patient information, to protect patient privacy.
We strongly believe in the potential for artificial intelligence toenhance, but never replace physician decision-making. Following this principle, the model must be user-friendly and provide clinicians with features they want and need, so we must study the aifred solution’s integration into clinical workflow and any effects on clinician efficiency and doctor-patient interaction.
Open Label Trials
Safety and effectiveness of the model must be assessed in open-label trials where both clinicians and patients know when our model is being used. A group of physicians using our model will be compared to a group practicing usual care, and patient outcomes will be compared between the two.
Randomized Control Trial
After open label studies, we will conduct one or more randomized control trials, testing our model against a “dummy” model and against a “practice as usual” group. This will help us determine how efficacious the aifred solution is.