Applications Of Machine Learning For Psychiatry Joseph Geraci PhD Molecular Medicine, Queen’s University
Dis isclo losure This presentation (the “Presentation”) does not constitute an offer to sell or a solicitation of an offer to buy any security and may not be relied upon in connection with the purchase or sale of any security. Any such offer would be made only by means of a formal offering memorandum. No such offer or solicitation will be made prior to the delivery of a confidential investment memorandum, private placement memorandum, or similar offering documents (“Offering Documents”) . Offers and sales will be made only in accordance with applicable securities laws and pursuant to the Offering Documents, the shareholders agreement, subscription agreements and other definitive documentation. Projections and other forward‐looking information contained in this Presentation, including all statements of opinion and/or belief, are based on a variety of estimates and assumptions, including, among others, market analysis, estimates and similar information. These estimates and assumptions are inherently uncertain and are subject to numerous business, industry, market, regulatory, competitive and financial risks. There can be no assurance that the assumptions made in connection with the projections will prove accurate, and actual results may differ materially, including the possibility that an investor may lose some or all of its invested capital. Dr. Joseph Geraci holds significant shares in companies that hold machine learning IP created by him and he also holds an academic position at Queen’s University in Ontario as a professor of molecular medicine. NetraMark, a holding parent company, provided financial support for Dr. Geraci to be at the ISCTM meeting. 2
Irreducible errors, bias, and variance? Explainability? The only hard thing you need to learn about machine learning today! • Irreducible Errors: sometimes the data will just not let the machine learn an aspect from your data. Some data is just…garbage…you know the saying • Bias: the model is underfit = it is not using all the data it has access to in order to create the model. Example: this vehicle is a “bus” because it is yellow and large • Variance: The model is overfit = the model is inflexible and only understands the data it was trained on. Example: A model that ”learns” that people with the name Jennifer, who are tall, drink tea instead of coffee, studied on the East coast, and do not drive, earn more money due to a biased data set. • Explainability: an algorithm that can reveal precisely what variables were used to make a decision and how it did it is a highly explainable system
Commonly Used Machine Learning Methods High Bias + Explainable • Logistic Regression Linear Regression
Commonly Used Machine May have high Variance + Explainability may be low. Learning Methods • Support Vector Machine Example: How do we separate these two groups with a simple linear boundary? By Larhmam - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=73710028
Commonly Used Machine Learning Methods – Support Vector Machine continued… • Add a dimension + transform the space and enjoy the freedom!
Commonly Used Machine Learning Methods • Tree based methods – e.g., Random Forest & XGBoost These methods attempts to tackle bias and variance in different ways and they are quite powerful at reducing errors. Amazing methods for learning from numerical data. Can be difficult to Explain.
Commonly Used Machine Learning Methods Capable of • Neural Networks great complexity so large data will help reduce Variance issues here – Not very Explainable but methods are emerging Supervised Learning VS Reinforcement? GANs? One AI tries to fool the other
What Is Augmented Intelligence? • This is a term some of us have been using for years to describe the process that results when one uses machine learning as an exploration tool – it is obvious and not complicated • The term is meant to highlight situations where the algorithms can learn about things that a human user does not know and where a human can interact and improve what the machine learns – True Co Co-Operatio ion
The difference between ordinary machine What they learning and Augmented Intelligence: Note the additional information in terms see… of subpopulations Note that in the AugAI paradigm clinicians and scientists can interact with the revealed subpopulations as to what is driving the different disease states and understand better how to treat them
Bip ipolar Dis isorder: : A New Taxonomy For r Subtypes Via ia Augmented In Intelligence – DEMO
How Can These Methods Be Used In A Psychiatric Clinical Trial? Examples From Commercial Use Cases • Placebo response prediction – Clinical trial optimization via an understanding of placebo response • Adverse Event (AE) detection – During a trial Augmented Intelligence can • Help identify factors that may predispose to the AE • Help identify subpopulations at particular risk for the AE • Subpopulation discovery – A new taxonomy of Psychiatric Disorders
Placebo Response model DeepCrush CAN PREDICT Placebo Response • The top right hand corner consists of an almost pure group of placebo non-responders to a drug (bipolar-depression) • This map was purely based on the patients’ “attitude” towards “medication” and their desire to get better • For example one of the questions was a measure of “TO WHAT DEGREE THEY WISHED TO GET BETTER”
Hig igh Dim imensional Anomaly Detection • Unsupervised, high dimensional anomaly detection can be used to ‘tip off’ clinical scientists to: • Patients that may not respond • Patients that may react adversely • A subgroup for which the drug works particularly well Example: Placebo Super-Responders – the two circled patients below responded “too Placebo Super-Responders well” and were discovered by the machine with no supervision. The machine aided the trial by corroborating the suspicion and revealing the pertinent variables
Hig igh Dim imension Anomaly Detection • Example 2: A trial had an adverse reaction event and they asked us to understand if the machine was able to elucidate – it not only had this patient as a ‘sub - outlier’, but there was another individual with a very similar profile and the driving variable was strongly aligned with the adverse event that occurred. Zoomed in perspective of a drug responder sub- population for a psychiatric group. By finding the patient on this map that had an adverse reaction we Adverse Reaction were able to identify Patients another who may have been at risk.
How Can Pharma Use These Tools? Research Development Commercial Point of care (Pre-clinical) (Trials) 1. Precision Targets & 5. Precision 2. Drug Resurrection 4. Drug Repurposing Disease Definition Treatments* 3. Clinical Trial Optimization Augmented Intelligence provides a road to the RDoC vision of establishing a new taxonomy of psychiatric disorders by integrating biology with established clinical scales. It does this as the machine is able to provide insights back to the human and together, the taxonomy will reveal itself. 16
The Future: Quantum ML Quantum Ready Technology Available now A path to the future of machine learning Biological Networks Quantum Computation In The Near Future Liver and Brain Organoids 17
THANK YOU Joseph Geraci geracij@queensu.ca
Recommend
More recommend
