The Risks of Medical Cannabis, Counted Honestly: Dependence, Hyperemesis, Heart, Psychosis, Cognition, Driving, and Drug Interactions

Two numbers set the frame. In a meta-analysis of 21 epidemiological studies published in Addictive Behaviors, 22 percent of people who had used cannabis met criteria for a cannabis use disorder, and among young people using weekly or daily the risk of cannabis dependence reached 33 percent. A 2026 JAMA review put the figure among people using cannabis specifically for medical purposes at 29 percent.

That is not a fringe complication. It is the single most common adverse outcome in cannabis medicine, and it is the one least likely to be raised at the point of sale.

Patients are entitled to a real risk accounting before they start, not after something goes wrong. A physician who describes cannabis as safe without qualification is making a claim the literature does not support, and a physician who describes it as dangerous without numbers is making a claim the literature does not support either.

The practical value of quantifying risk is that it becomes possible to reduce it. Nearly every serious signal on this page tracks with dose, potency, frequency, route, and the medications a patient is already taking. Those are modifiable.

Janni Leung and colleagues at the University of Queensland pooled 21 epidemiological studies published between 2009 and 2019 in Addictive Behaviors. Among people in the general population who had used cannabis, 22 percent (95 percent CI 18 to 26) met criteria for a cannabis use disorder, 13 percent for cannabis abuse, and 13 percent for cannabis dependence. In cohort studies restricted to young people using weekly or daily, the risk of developing cannabis dependence rose to 33 percent (22 to 44).

The 2026 JAMA review of therapeutic cannabis by Hsu, Shah, Jordan, Gold, and Hill reports a pooled figure of 29 percent for cannabis use disorder among people using cannabis for medical purposes specifically. Medical framing does not confer protection.

Population-level data confirm the trend is moving the wrong way. Analysis of 186,823 adults in the 2021 to 2024 National Survey on Drug Use and Health found past-year cannabis use disorder rising from 5.9 percent to 7.4 percent of all adults, with increases across every sociodemographic subgroup examined.

Withdrawal is the tangible face of this. A meta-analysis of 47 studies and 23,518 participants in JAMA Network Open found a pooled cannabis withdrawal syndrome prevalence of 47 percent among people with regular or dependent use, ranging from 17 percent in population samples to 87 percent among inpatients. Irritability, sleep disruption, appetite loss, and restlessness on stopping are not evidence of weak character. They are a described syndrome.

Textbooks and review articles described cannabinoid hyperemesis syndrome as rare for years without any prevalence data behind the word. Joseph Habboushe and colleagues at Bellevue Hospital in New York went and measured it. Among 155 emergency department patients aged 18 to 49 who reported smoking cannabis at least 20 days per month, 32.9 percent (95 percent CI 25.5 to 40.3) met the study’s criteria for having experienced the syndrome, defined by cyclic nausea and vomiting relieved by hot showers.

Extrapolated, the authors estimated roughly 2.75 million Americans may experience a phenomenon consistent with hyperemesis annually. The extrapolation deserves caution, because the sample was a convenience sample in one urban public hospital and the case definition was symptom-based rather than diagnostic. The direction of the finding is still hard to dismiss.

The clinical trap is that hyperemesis presents as a gastrointestinal problem and patients often increase cannabis use to treat the nausea, which worsens it. Compulsive hot bathing is the giveaway. The only reliably effective treatment is cessation. A 2024 systematic review in Academic Emergency Medicine examined capsaicin cream and dopamine antagonists across seven studies and 492 patients, finding mixed evidence for capsaicin and possible benefit from haloperidol and droperidol, with the authors calling for rigorous trials.

A 2025 systematic review and meta-analysis in Heart by Storck and colleagues pooled 24 pharmacoepidemiological studies and reported a risk ratio of 1.29 (95 percent CI 1.05 to 1.59) for acute coronary syndrome, 1.20 (1.13 to 1.26) for stroke, and 2.10 (1.29 to 3.42) for cardiovascular death in cannabis users. Restricting the analysis to cohort studies gave a comparable estimate of 1.32 (1.01 to 1.73).

The 2026 JAMA review reports absolute figures that are easier to weigh at the bedside. Comparing daily inhaled cannabis use with nondaily use: coronary heart disease 2.0 percent versus 0.9 percent, myocardial infarction 1.7 percent versus 1.3 percent, and stroke 2.6 percent versus 1.0 percent.

Most of this evidence is observational and most of it involves smoked cannabis, so confounding by tobacco co-use is a persistent limitation the authors acknowledge. What the data support is a specific clinical position rather than a blanket warning: in a patient with established coronary disease, poorly controlled hypertension, or a prior stroke, inhaled cannabis is a harder recommendation to justify than a low-dose oral preparation, and the acute tachycardia and orthostatic effects of THC are a separate practical concern in older patients.

The psychosis literature is often reported as though cannabis either causes schizophrenia or does not. The more accurate reading is that risk scales steeply with frequency and THC concentration.

The EU-GEI multicentre case-control study, published by Marta Di Forti and colleagues in Lancet Psychiatry, compared 901 patients presenting with first-episode psychosis across 11 European sites and Brazil against 1,237 population controls. Daily cannabis use carried an adjusted odds ratio of 3.2 (2.2 to 4.1) versus never use, rising to 4.8 (2.5 to 6.3) for daily use of high-potency cannabis defined as 10 percent THC or more. Assuming causality, the authors calculated that removing high-potency cannabis would prevent 12.2 percent of first-episode psychosis cases across all sites, 30.3 percent in London, and 50.3 percent in Amsterdam.

A separate dose-response meta-analysis by Marconi and colleagues in Schizophrenia Bulletin, covering 66,816 individuals across 10 studies, found an odds ratio of 3.90 (2.84 to 5.34) for psychotic outcomes in the heaviest users compared with nonusers.

The Ontario cohort work by Daniel Myran and colleagues in JAMA Psychiatry adds the piece most relevant to a clinic. Among 9.8 million people with no history of psychosis, an emergency department visit for cannabis-induced psychosis carried an adjusted hazard ratio of 241.6 for later transition to a schizophrenia spectrum disorder. An emergency visit for cannabis use without psychosis still carried a hazard ratio of 14.3, and because those visits are far more common they produced a greater absolute number of transitions.

None of this establishes causation in an individual. It does mean that a personal or family history of psychosis, and daily use of high-potency product in an adolescent or young adult, are the two situations where the risk calculus changes most sharply.

The cognitive evidence is frequently overstated in both directions. J. Cobb Scott and colleagues at the University of Pennsylvania meta-analyzed 69 studies covering 2,152 cannabis users and 6,575 comparison participants in JAMA Psychiatry. Frequent or heavy cannabis use was associated with a small overall reduction in cognitive functioning, mean d of -0.25 (95 percent CI -0.32 to -0.17). Critically, in the 15 studies requiring more than 72 hours of abstinence, the effect was -0.08 and no longer statistically distinguishable from zero.

The honest summary is that measurable cognitive effects in young people are small on average and substantially reflect residual acute effects and withdrawal rather than fixed injury. That is reassuring compared with what patients are often told, and it is not a statement that heavy adolescent use is harmless.

Driving is the risk that shows up in real life. Rogeberg and Elvik reanalyzed and extended the crash literature in Addiction, pooling 28 estimates from 21 observational studies with a combined sample of roughly 240,000, and found an odds ratio of 1.36 (1.15 to 1.61) for collision after acute cannabis use, falling to 1.22 (1.10 to 1.36) in meta-regression. That is a genuine increase and it is smaller than earlier published estimates, which the authors showed had been inflated by methodological problems.

For timing, McCartney and colleagues reviewed 80 publications and 1,534 outcomes in Neuroscience and Biobehavioral Reviews and predicted that most driving-related cognitive skills recover within about five hours of inhaling 20 mg of THC, with almost all recovering by about seven hours. Oral dosing takes longer. Five hours after inhalation is a defensible minimum, not a generous one.

THC and CBD are metabolized by CYP3A4, CYP2C9, and CYP2C19, the same enzymes that handle a large share of prescription medicine. A 2024 systematic review in Frontiers in Pharmacology by Nachnani and colleagues screened 4,600 reports and identified 31 in which cannabinoids altered pharmacokinetics or produced adverse events, involving 16 narrow therapeutic index medications. Warfarin, valproate, tacrolimus, and sirolimus were the most frequently reported and, in the authors’ judgment, pose the greatest risk. In 18 of the 31 reports, or 58 percent, clinicians discovered an unexpected serum level of the prescribed drug.

The clearest documented interaction comes from epilepsy practice. Tyler Gaston and colleagues at the University of Alabama at Birmingham monitored serum antiseizure drug levels in 39 adults and 42 children during open-label pharmaceutical cannabidiol titration up to 50 mg per kilogram per day, reported in Epilepsia. Rising CBD dose produced significant increases in N-desmethylclobazam, topiramate, and rufinamide and a fall in clobazam, with sedation in adults tracking the N-desmethylclobazam level. Aspartate and alanine aminotransferase levels were significantly higher in participants taking concomitant valproate.

That hepatic signal appears in the registration trials as well. In the New England Journal of Medicine trial of cannabidiol in Dravet syndrome, abnormal liver function test results occurred more often on cannabidiol than on placebo, alongside diarrhea, vomiting, fatigue, pyrexia, and somnolence, and there were more withdrawals in the cannabidiol arm. Liver enzyme monitoring is part of prescribing pharmaceutical cannabidiol for exactly this reason.

The practical point for a consumer CBD user is different but not trivial. Retail CBD is sold in doses far below 20 mg per kilogram, so the hepatic signal from epilepsy dosing does not transfer directly. The interaction risk with narrow therapeutic index drugs does, because enzyme inhibition happens at lower exposures than hepatocellular injury.

The evidence quality varies sharply by risk, and treating it as uniform is the most common error on both sides of this argument. Cannabis use disorder and withdrawal rest on large meta-analyses of epidemiological data with consistent estimates across settings, which is strong. The psychosis association rests on well-conducted case-control and cohort work with a clear dose-response gradient, which is strong for association and still contested for causation. The cardiovascular estimates come almost entirely from observational studies in which tobacco co-use is difficult to separate out, which is moderate at best.

Cannabinoid hyperemesis has the weakest formal evidence base and one of the strongest clinical footprints. The prevalence figure quoted here comes from a single convenience sample with a symptom-based case definition. Emergency physicians in legal jurisdictions describe seeing it constantly. That mismatch between measured evidence and clinical experience is a reason to study it properly, not a reason to dismiss it.

Nearly all of this is observational, and people who use cannabis heavily differ from people who do not in ways that are hard to adjust away: tobacco, alcohol, other substances, socioeconomic position, and pre-existing psychiatric illness. Reverse causation is a live problem in the psychosis literature in particular, since early psychotic symptoms can prompt self-medication.

Exposure measurement is crude throughout. Studies classify people as users or nonusers, or count days of use, while the actual pharmacological exposure depends on THC concentration, route, inhalation technique, and tolerance. The EU-GEI study is unusual in separating high-potency from low-potency product, and it is precisely that separation that produced the largest effect estimates.

Publication and framing effects cut both ways. The Rogeberg and Elvik reanalysis is a useful corrective: two widely cited driving risk estimates turned out to lie outside the revised confidence interval once methodological problems were fixed. Numbers from this field should be treated as provisional.

None of these studies establish that medical cannabis caused a specific outcome in a specific patient. Meta-analyses of population data describe average associations across groups, and the absolute risks for most of these outcomes remain low for most people.

The literature also does not tell us the risk profile of the products most medical patients actually use. Most safety data come from smoked cannabis in recreational patterns. Low-dose oral preparations, high-CBD formulations, and supervised medical regimens are underrepresented, and it would be wrong to assume either that they carry the same risks or that they carry none.

Cannabis medicine has a credibility problem that comes from its own advocates rather than from its critics. When a field describes a compound as non-toxic and well tolerated while emergency departments fill with cyclic vomiting and roughly a fifth to a third of regular users meet criteria for a use disorder, it invites the dismissal it then complains about.

The opposite failure is equally common in mainstream medicine, where a single association is treated as a prohibition and patients hear nothing quantitative at all. Both failures cost the same thing: a patient who stops disclosing, and therefore stops being managed.

Every major signal here is dose-dependent, potency-dependent, or route-dependent. That is the argument for clinical involvement rather than against cannabis. A patient who is guided toward the lowest effective dose, an oral or low-temperature vaporized route, a characterized product, and a medication interaction review is carrying materially less risk than the same patient buying on their own.

I tell new patients the dependence number before they ask, because it is the risk they are most likely to meet and the one nobody at a dispensary counter will mention. Roughly a fifth to a third depending on which population you look at. Most patients hear that and are neither frightened nor dismissive. They adjust.

Hyperemesis is the one I look for hardest, because it hides. Someone comes in with months of cyclic vomiting, a negative gastrointestinal workup, and a cannabis habit they have been increasing to control the nausea. The hot shower question settles it faster than any test, and the treatment is the thing the patient least wants to hear.

What I do not do is hand a patient a list of risks and leave. The risks on this page are mostly modifiable. Lower the dose, change the route, avoid driving for the evening, review the warfarin, check the family psychiatric history before recommending anything high in THC to a nineteen-year-old. That is the work.

Medical cannabis carries a real and quantifiable set of risks. Dependence affects roughly 22 to 29 percent of users depending on the population, hyperemesis is far more common than its textbook description, the cardiovascular association is modest but consistent, the psychosis risk concentrates in daily high-potency use and in people with psychiatric vulnerability, driving impairment persists for about five hours after inhalation, and interactions with warfarin, valproate, tacrolimus, and sirolimus deserve active management. Every one of those is easier to manage with a physician involved than without.

Carry forward the effect sizes, not the adjectives. Cannabis is neither benign nor uniquely dangerous, and the useful question is never whether it has risks but which risks apply to which patient at which dose by which route. The single most transferable finding on this page is that almost every signal scales with frequency and potency, which means the clinical lever exists.