Causes and Associations in Single-Individual Analysis (CASIA) [pronounced: ka-sha]
The Situation: You have a lot of data from your wearable or implantable device, sensor, or mobile app. You have a recurring outcome you’d like to change (e.g., weight, irritable bowel syndrome, migraine headaches, asthma attacks, chronic pain, blood glucose levels). You’ve identified possible triggers, but their effects may take some time to appear---and it may be expensive or painful to test all or even just a few of them.
The Challenge: Design experiments to conduct on yourself to characterize the effects of only the most likely triggers.
I am the principal investigator (PI) of the CASIA Project. My goals are to establish the feasibility of applying causal inference methods to improve causal discovery for personalized/precision health, and to develop analysis and analytic methods based on n-of-1 randomized trials (N1RTs), for both observational and experimental studies. (An N1RT is often a randomized single-subject crossover trial, also called a "single-case design".) I was supported by a 2017-2018 Stanford Center for Clinical and Translational Research and Education (Spectrum) Pilot Grant for Population Health Sciences, titled Improving personalized medicine through n-of-1 causal inference and predictive modeling (N1CPM), with primary co-investigators Professor Lorene Nelson (co-PI) and Katherine Holsteen.
The CASIA Project develops the theory and application of statistics for studying one individual's recurrent characteristics and patterns. I call this field esametrics (pronounced "EE-sa-metrics"), from isa (pronounced "ee-SA"), the Filipino/Tagalog word for "one". I borrowed this concept from the psychology term "idiographic", which refers to studies focused on a single person.
To meet The Challenge above, my general approach is to analyze time series partitioned into periods akin to those of an N1RT. The resulting analyses rely on the potential-outcomes (i.e., counterfactual) framework to draw causal inference---strengthened using machine learning---from these observational data, in what I call an n-of-1 observational study, with the target estimand being what I call an average period treatment effect (APTE; Daza, 2018). Published examples of estimated APTEs include an APTE of physical activity on weight (Daza, 2018), and an APTE of sleep deprivation on continuously monitored blood glucose level (Daza et al, 2020). Further development of APTE-based analysis can be found in Daza (2019).
General applications (i.e., not specific to health) of the N1CPM methods include situations described by the following criteria (modeled on Karkar et al, 2015):
Recurrence (Outcome): The outcome you want to change occurs regularly.
Precedence (Exposure): The exposure that might change the outcome precedes said outcome.
Classifiability (Exposure): The exposure can be categorized into generally non-overlapping treatment or intervention levels (e.g., treatment vs. control, A vs. B).
Manipulability (Intervention): You must be able to manipulate or otherwise change the intervention levels.
Dynamism (Effect): The effect of the intervention on the outcome may vary over time, and may take time to stabilize (if at all).
Prof. Nelson and Dr. Holsteen led a larger study called Studying TRiggers in Everyday Activity for Migraine (STREAM), where we hope to apply these N1CPM methods to help each participant discover what triggers their migraine headaches, and how they can change or avoid them. We foresee the application of these methods to many other recurring chronic conditions, such as asthma, functional gastrointestinal disorders (e.g., irritable bowel syndrome), and chronic pain.
- Daza EJ. Causal analysis of self-tracked time series data using a counterfactual framework for n-of-1 trials. Methods of Information in Medicine. 2018 Feb;57(01):e10-21.
Karkar R, Zia J, Vilardaga R, Mishra SR, Fogarty J, Munson SA, Kientz JA. A framework for self-experimentation in personalized health. Journal of the American Medical Informatics Association. 2015 Dec 7;23(3):440-8.