Hand-crafted Bayes Net Model of Individual Aging

We have “handcrafted” an initial BayesExpert model for our Rejuve.AI app to give users information relevant to their longevity while the network of scientist contributors is forming.¹⁹

This “hand-crafted” literature-based Bayes network combines many clinical studies related to the nine plus one hallmarks of aging with actions users can take to improve their chances of preventing age acceleration due to each of the hallmarks. The Pomegranate Bayesian net that Bayes Expert generates from hand coded rules has 315 nodes.²⁰ Of those, 162 are discrete distributions , the “leaves” of the Bayesian network, that derive their priors from 85585 NHANES data contributors who took blood tests.²¹ The remaining 153 nodes are conditional probability tables, each created from the rules. 63 of these were created from dependency rules, and 90 are logic rules using and, or, and avg to organize the data. The relations in the dependency rules come from 62 meta analyses and systematic reviews from reputable medical journals, based on 1500 gold standard Randomized Controlled Trials, on a total of 12.5 million subjects. Their priors also come from the NHANES data.

Users currently enter health data via surveys that are very similar to NHANES questions, along with syncing of wearable health tracker devices. We collect 13 wearable device signals, which we use to prefill survey responses in the app and send to an ADTK anomaly detector. The anomaly detector uses medical literature-based thresholds to detect autoregression, interquartile range, and level shift anomalies in the signals.²² Signal anomalies are sent to the Bayesian Network, which interprets the survey data using rules to determine what aspects of the data entered most helped and harmed the hallmarks of aging risk scores. The user is then sent both scientific and user-friendly literature about the most important things we know about what they are doing and that appear reasonably likely to have the potential to change the risk to the hallmark, as inferred by the Bayesian network.

Each study relation (such as relative risk) in the network has a validation score that indicates how much the probability value window had to change in order to have a feasible result in quadratic programming, where a feasible result indicates a data match. The present net has an average validation score of 0.056, but a median of 0.008 and a standard deviation of 0.099. The score indicates that the probability had to change by an average of 5%, which is a small amount. The GCN's fitness function will also use this validation score to find more consonant sets and conditionals that make them consonant, that is, with a lower validation score, automatically. However, the median is less than 1%, indicating a good overall match. The GCN will investigate the circumstances under which cliques of agreeing crowdsourced studies help solve problems more effectively than other cliques of studies.

¹⁹ Rejuve.AI (2022) Longevity Bayes. Available at: https://github.com/Rejuve/bayesnet/blob/master/sn_bayes/longevity_bayes.py (Accessed 5 October 2022) ²⁰ Schreiber, J. (2018) Bayesian Networks. Available at: https://pomegranate.readthedocs.io/en/latest/BayesianNetwork.html (Accessed 5 October 2022) ²¹ National Center for Health Statistics (2022) National Health and Nutrition Examination Survey. Available at: https://www.cdc.gov/nchs/nhanes/index.htm (Accessed 5 October 2022) ²² Arundo Analytics (2020) Anomaly Detection Toolkit. Available at: https://adtk.readthedocs.io/en/stable/ (Accessed 5 October 2022)