Building a Learning Marijuana Surveillance System.

Colorado,Washington,andAlaskahave legalizedmarijuana, thus allowing a commercial industry to develop andmarketaproduct thathasbeen, ineffect, illegal since 1937. Commercialization of other addictive substances such as alcohol and tobacco created a cascade of negativepublichealth consequences, leading to reactiveand ineffective attempts at regulation of amature industry. Although there are calls to study the effects of marijuana legalization,1 surveydatahavegeneratedmixedresults thus far. For example, Stolzenberg et al2 reported amplified adolescent use in states that legalized marijuana, while Hasin et al3 found no significant changes. These conflicting results highlight the need for a learning marijuana surveillance system that uses varied, robust, and real-time inputs to connect use rates, acute harms, andmore subtle prediagnostic indictors ofmorbidity to changes in policy and product. Whileoftenportrayedas “safe,”marijuanause is associated with morbidity and mortality. Altered sensorium may be experienced as pleasurable, but also increases reaction timeand riskof car crashes,particularly among inexperienced drivers. Exposure to highly concentrated marijuana products can result in acute psychotic reactions, which can in turn result in suicide or other injuries. Systems for tracking acute harms rely on reports from a network of hospital emergency departments thatmaynot connect clinical presentationswith changes in product potency, formulation, or use patterns. Forexample, anuptick in theuseof synthetic cannabinoids createdapublic health crisis, as thousandsof users experienced severe adverse reactions, a pattern that health officials were slow to recognize.4 This illustratesourpoorability tospotandrespondtoacuteharms from shifting drug epidemics using population survey data and case reports. Acute harms make up only a small fraction of consequences from marijuana use. The real challenge of learning surveillance lies in detecting effects of longterm use. Repeated exposure tomarijuana during critical windows of brain development is associated with marked anatomical changes,5 poor functional outcomes,6 and seriousmental healthdisorders.7Marijuana surveillance systems are poorly configured todetect these downstream sequelae, relying primarily on largenationaldatasetsthattrackself-reporteduse.While effective for detecting large trends, these systems are relatively insensitive to fine shifts in use that may unfold in response to state or local policy change. They do not distinguish between single-occasion and frequent past-monthuse that falls belowadaily threshold,nordo they account for potency, product type, or quantity of consumption. These factors define dose and therefore outcomes, which vary with state-level policies.8 A learning marijuana surveillance system9 is needed to augment traditional survey surveillance with novel approaches such as mining social media content and crowdsourcing product reports. By inferring information from public communications or by engaging social media users who intentionally share information, an augmented system would significantly expand the volume of information about use, fostering high-resolution views into demographic, temporal, and spatial trends. Using a model that is flexible and designed for continuing improvement, this system could extend across social sectors to enable quantification of school, workplace, and interpersonal outcomes not currently evaluated. These approaches could illustrate marijuana use patterns in relation to other health behaviors (eg, use while driving), events (eg, parties, holidays, or birthdates), or harms (eg, missing class, failing examinations, or unintentional injuries that do not rise above threshold for emergency treatment). Content analysis of communications could shed light on the diffusion of opinions and beliefs across social networks, elucidating both problems and sentiment and allowing policy makers to act from evidence while remaining sensitive to popular convictions. Such an approach is revolutionizing infectious disease surveillance and has alerted the public health community to shifting sentiment about diverse health issues such as fluoridation and vaccination.10 Better surveillance would also prepare health care professionals to respond to patient questions and address safety concerns. Where data are sufficiently granular, findings from a learning marijuana surveillance system could alert researchers and policy makers to shifts in consumption and associate them with regional and possibly even local interventions, enabling policy course correction. A learningmarijuana surveillance systemwouldengagestakeholdersacrosssocial sectors.Partnershipwith educational systems could enable investigation of traditionally collected reports of school outcomes in relation to reports of marijuana use trends. This might revealproblemsstemmingfromchronicmarijuanausethat are not typically quantified or factored into policy decisionsabout legalizationandavailability. If executedwell, an expanded marijuana surveillance system would be comprehensive, tuned to detect harms from recreational as well as unregulated medical use, a gap area. As themarketunfolds, informationaboutmentalhealth, medical, or school outcomesmight suggest we impose restrictions on maximal tetrahydrocannabinol content or bans on certain products that are particularly attractive to youth. With an accurate assessment of downstream consequences in hand, communities may decide to forgo legalization evenwhere state law permits it, as has occurred in Colorado. Commercialized marijuana is a rapidly moving target. High-potencymarijuana plants and products, new VIEWPOINT