Did a Smartphone Treatment Reduce Methamphetamine Use? What a 100-Person Trial Actually Measured

Short answer: An eight-week randomized study found significant within-group reductions in reported cue-induced craving among some smartphone-program participants, with a model-dependent improvement in visual recognition. It did not establish a reduction in methamphetamine consumption using toxicology, nor did the published paired analyses alone prove an incremental treatment-by-time effect over telephone counseling. All 100 enrollees were reported without positive toxicology screening in the relevant observation interval, so screening was excluded as an outcome. No participant-specific venous whole-blood methamphetamine concentration or real-world impairment endpoint was measured.

The genuine original empirical source is Zhang and colleagues (2022), Preliminary efficacy of a digital therapeutics smartphone application for methamphetamine use disorder: An experimental study, Frontiers in Psychiatry, DOI 10.3389/fpsyt.2022.1027695. The study illustrates why a reported craving score, a statistically imputed memory outcome, an engagement log, and a chemical drug measurement must not be treated as equivalent forms of evidence.

What did the randomized study actually establish?

The investigators studied a bundle of behavioral services delivered by a smartphone application, not a medicine capable of directly removing methamphetamine from blood. One hundred adults entering community rehabilitation were randomized to eight weeks of the digital program or usual telephone-based support. In the digital group, some self-reported craving scores fell significantly between baseline and week eight, and a visual-recognition task improved in the analysis that filled in missing observations statistically. Those are narrower findings than a demonstrated reduction in methamphetamine consumption. The original study did not use toxicology results as the efficacy endpoint because no participant had a positive screening result during the stated observation interval.

The distinction is especially important when the paper is used in a forensic setting. An application can change what a person reports while viewing drug-related pictures; the same experiment can record memory-test scores and app logins. Neither kind of record quantifies methamphetamine or amphetamine in legally collected venous whole blood. No named individual's dose, time of ingestion, impairment or act at a particular event was reconstructed in this study. The paper is promising early behavioral-intervention research, but it is not a pharmacokinetic or impairment-validation study.

Who were the 100 independently randomized participants?

Between January and March 2021, four community rehabilitation centers in Chengdu, China, recruited adults diagnosed with methamphetamine dependence under DSM-IV criteria. Eligible participants were 18 to 50 years old, able to operate an Android or iOS smartphone, and free of additional diagnosed mental health conditions under the study's exclusion rule. Fifty-two people were assigned to digital therapeutics, or DTx, and forty-eight to treatment as usual, or TAU. The participants were entering community care rather than being given a controlled recreational-drug challenge. Their historical methamphetamine exposure was not an investigator-administered dose.

The unit of randomized allocation was the person. The research contains numerous phone taps, app visits, picture presentations, practice trials and follow-up responses, but those activities remain nested within the same one hundred people. Forty visits by one individual would not constitute forty independent clinical cases. The two arms also differed in their access to the app and the manner of assessment; these differences matter when interpreting observed self-report and the probability that a follow-up questionnaire would be completed.

What exactly was the smartphone intervention?

The application, named WonderLab Harbor, combined four elements usually studied separately: interactive cognitive-behavioral therapy (ICBT), approach-bias modification (ApBM), game-based visual memory training, and contingency-management-style rewards. The ICBT modules addressed recognizing craving triggers, coping strategies, refusal skills, problem solving, thoughts about drugs, and related health topics. Sessions unlocked sequentially, while already unlocked material could be reviewed. The intervention therefore involved psychoeducation and repeated structured choices rather than administration of a stimulant, sedative or anti-craving compound.

The bias-training component used pictures of methamphetamine-associated items and alternative healthy activities. A participant swiped in response to image orientation, with a designed tendency to push drug-related images away and draw neutral or positive images closer. Each session comprised 156 image presentations, with repeated presentations of twelve underlying cue pictures. The investigators recorded responses and response times. Those 156 trials are technical or behavioral observations nested within a participant; they are not independent proof that the participant later refused an actual offer of methamphetamine.

The visual memory game displayed shapes briefly, removed them, sometimes rotated the display, and asked the user to reconstruct the correct positions. Difficulty varied with the number of symbols and grid complexity. Rewards, daily check-ins and phone-credit redemption were intended to reinforce adherence. Each 1,000 points could be exchanged for a small mobile-phone credit under the program's rules. Such incentives are part of the studied treatment package, not evidence that the same cognitive outcomes would result from a bare app without contact, rewards, or expectancy.

What did the control arm receive?

The forty-eight TAU participants received conventional community-based counseling. Social workers and therapists contacted them by telephone, partly because COVID-19 control measures impeded ordinary in-person counseling. The calls addressed work, family, stress and craving suppression, and the reported mean phone contact was about eleven minutes. The comparison therefore was not a digital placebo matched for daily engagement or contact time. The digital arm was offered multiple self-directed activities and incentives, while the control received episodic telephone support.

Randomization is a valuable starting point, but it does not itself blind participants or equalize all routes of data collection. Digital participants were instructed to install the study app on their own phones and complete picture-based craving assessments there. Controls generally used a social worker's phone to complete the same type of assessment at the center. Differences in setting, convenience, privacy, response rate and investigator contact could influence who completed assessments and how scores were recorded. The original study's paired statistical tests also address change within each arm, a different question from a formal between-arm interaction.

What was the actual primary outcome?

The primary outcome was cue-induced craving, not biologically demonstrated methamphetamine use. Participants saw thirty methamphetamine-associated images and rated craving for each on an eleven-point visual-analog scale from zero to ten; the thirty ratings were summed. A high total represented stronger reported response to that particular image set. The application also displayed thirty healthy-lifestyle images, but the stated drug-craving score was the sum across drug-related cues. Consequently, changing the score could indicate less self-reported cue reactivity in that assessment context, not necessarily a reduction in the frequency, amount or recency of illicit drug consumption.

The distinction between psychological craving and biochemical exposure applies even if the same respondent provides reliable answers. An individual might report less craving while remaining at risk of use, or might refrain from use without exhibiting a large questionnaire change. The study did not establish sensitivity or specificity of this image score for recent methamphetamine consumption, and did not calibrate craving units to a whole-blood concentration or an impairment threshold. It therefore should not be described as a validated diagnostic test for current intoxication.

Why did the authors omit toxicology as an outcome?

The Methods section expressly explains that toxicology screening was not used as an outcome because community screening was predominantly negative: none of the one hundred study participants had a positive toxicology screening result during January through June 2021. With no positive outcomes to distinguish one arm from the other, the investigators selected cue-related self-report and cognitive assessment rather than urine-test change. That is not the same as demonstrating that the app achieved abstinence. The negative screening history preceded or overlapped the intervention and provides no discriminating randomized efficacy endpoint.

The original article does not provide participant-level confirmatory chromatograms identifying individual methamphetamine and amphetamine peaks, nor an assay-validated concentration-time series in urine, separated plasma, serum or legally collected human venous whole blood. A screening status is different from quantitative identification of parent drug or metabolite. If a toxicology laboratory had validated screening information, the tested specimen, specific cutoff, collection time, detection window and confirmatory method would all be needed before using it for individual forensic inference. The study does not furnish that independent analytical record.

Why are response rates central to the interpretation?

Follow-up was highly unequal across the arms. The original table reports baseline craving response rates of 46 of 52 digital participants, about 88%, compared with 22 of 48 controls, about 46%. At the eight-week assessment, the corresponding numbers were 31 of 52, about 60%, and 16 of 48, about 33%. These are not trivial missing cells in an otherwise complete experiment. They mean the observed craving outcomes in each arm came from different proportions of the people originally randomized. The true biological denominator remains one hundred; the available complete-case outcome denominator is substantially smaller.

Selection may occur if people who feel better, use their phones more frequently or find a task tolerable are more likely to provide a follow-up. Conversely, some participants with worse symptoms might seek more app contact. Without observing the missing outcomes, the direction and magnitude of this selection are uncertain. A statistically significant paired change among respondents cannot simply be applied to every randomized participant. That is why the authors reported complete-case estimates and a separate multiple-imputation analysis rather than claiming that the observed respondents constituted the entire randomized trial.

Evidence table: allocation, assessment completeness and endpoints

The first evidence table reorganizes the study's original counts and directly observed scores by their proper denominators. Questionnaire completion rates measure how much outcome data were available, not the rate of relapse or the efficacy of therapy. The visual memory scores are task performance and should not be described as a general intelligence test. Reported baseline and post-intervention craving means were highly variable, emphasizing the limits of simple mean comparisons in a skewed sample.

Did craving fall significantly in the digital group?

In the participants with paired observations, the digital arm showed a Wilcoxon signed-rank effect size of −0.402 for cue craving, with reported p = .026. The associated bootstrapped point estimate was −0.396, 95% confidence interval −0.638 to −0.087, with bootstrap p = .011. Both expressions concerned the change within the app group over time, not a randomized treatment-by-time contrast between all fifty-two digital and forty-eight TAU participants. No author-supplied validated test converted the signed-rank estimate into a percentage reduction in real methamphetamine ingestion.

For the imputed dataset, Table 2 reported the digital group's craving estimate as −0.267, 95% confidence interval −0.435 to −0.099, p = .002. Under that modeling approach the within-group direction remained favorable. The placebo-like comparator, however, was not a matched sham app and had much more missing craving data. The reported analyses do not demonstrate that an identical measured craving outcome changed more under digital treatment than under the randomized control by a formally reported arm-by-time interaction. Positive within-group results and proved incremental efficacy must be kept conceptually separate.

Did a memory task improve, and how robust was that finding?

The Meaningless Figure Recognition Test presented novel geometric-like shapes for brief study and then asked participants to recognize them among distractors. It is a visual-recognition task, not a measurement of everyday memory, driving attention, criminal intent or detection of drug use. The complete-case digital-arm estimate was +0.312, with p = .094 and a bootstrap 95% interval from −0.047 to +0.631. Because that interval included zero and the p-value exceeded .05, the source did not demonstrate a statistically significant within-group memory change in the complete-case comparison.

After multivariate imputation by chained equations, or MICE, the reported digital-arm estimate was +0.220, 95% confidence interval +0.009 to +0.432, p = .041. This crosses the conventional statistical decision threshold, but is sensitive to how missing records are imputed. It is not a direct measurement of an unobserved person's cognition. Multiple imputation can be useful under justified missing-data assumptions; it cannot independently verify the true values of every absent response. The source therefore supports cautious discussion of a model-dependent improvement in this specific test, not a generalized claim of restored cognitive capacity.

Which emotional and comparator endpoints remained nonsignificant?

The app-arm analysis did not show a measurable change in its PHQ-9 depressive-symptom score or GAD-7 anxiety score. Under MICE the depression estimate was −0.005 with p = .967, and the anxiety estimate was −0.011 with p = .930. Both are genuine tested null findings rather than unmeasured outcomes. In the telephone counseling arm, the imputed within-group craving estimate was −0.158 with p = .347 and the visual-memory estimate was +0.012 with p = .920; depression and anxiety also did not show significant changes.

These distinctions prevent two opposing overstatements. It would be inaccurate to say the application significantly improved every outcome, because several measured endpoints did not reach significance. It would also be incorrect to conclude that the comparator offered no psychological benefit, because the study was not powered and analyzed to establish equivalence, and incomplete responses greatly limited the estimates. Nonsignificance is not proof of zero biological or clinical effect. The particular endpoint, instrument, contrast and sample should accompany every reported p-value.

Evidence table: original paired effect estimates and null findings

The second table presents the numerical within-arm estimates as reported in Table 2, with the missing-data method identified rather than treating imputed observations as newly recruited people. The values are nonparametric effect-size estimates and p-values, not measured plasma stimulant concentrations or relative risks of relapse. The distinction between an analysis within a randomized arm and a direct comparison of randomized arms is essential.

What is the difference between within-arm significance and a treatment effect?

Suppose a group receiving an app improves its craving score between baseline and the end of treatment. Several processes can explain some part of that change: the app's active exercises, counseling received outside the app, natural recovery during a rehabilitation program, regression to the mean, changes in living circumstances, or simply completing the same questionnaire a second time. Randomization supplies a control group so investigators can compare those competing explanations. The directly relevant efficacy question is whether the app-arm change differs reliably from the concurrent control-arm change, using a prespecified between-arm analysis.

The source instead emphasizes paired Wilcoxon signed-rank tests within each group, reported separately. One arm reaching p = .002 while another reaches p = .347 does not itself establish that the arm-to-arm difference is statistically significant. Comparing significance labels rather than testing their difference is a known inferential mistake. The sample also had unequal craving-response rates and different phone-based assessment arrangements. The results therefore support a preliminary signal and a feasibility discussion, but not the stronger assertion that the app caused a verified reduction in methamphetamine use relative to usual care.

How were missing data handled statistically?

Investigators reported conventional complete-case analyses and bootstrapped multiple imputation using chained equations. Complete-case analysis uses only people with the required baseline and post-intervention responses. When response is related to symptom severity, app engagement or other unobserved factors, that subset may not represent everyone randomized. Imputation instead uses patterns in observed variables to estimate plausible missing values, then combines results according to the method's assumptions. Neither strategy makes missing information physically measured or eliminates systematic nonresponse.

For the digital-arm visual-recognition task, interpretation changes visibly across those methods: p = .094 under complete cases versus p = .041 with MICE. That difference does not mean the actual participants changed their cognitive ability between calculations. It shows that the inferential result depends on missing-data assumptions and the finite set of observed responses. For forensic discussion, presenting both estimates is preferable to selecting the favorable significance result and omitting the other. Analyses of a person's real-world functional status require case-specific, contemporaneous measurements; a group imputation is not such a measurement.

Did the application retain engagement after eight weeks?

Usage logs provide an important feasibility measure separate from efficacy. By the end of the designed eight-week treatment period, six of fifty-two digital-arm users, 11.5%, had stopped using the application entirely. The study also reported module-specific cessation: eighteen, or 34.6%, had stopped the interactive CBT component, and fourteen, or 26.9%, had stopped the cognitive-training module. Some people continued logging in while using fewer of the assigned therapeutic exercises. An app icon remaining installed cannot be treated as evidence that a patient repeatedly practiced every component.

The researchers followed user activity for forty weeks, even though the planned clinical intervention and clinical endpoint assessment covered only eight weeks. The average user had logged in approximately 16.37 separate days by week eight and completed 4.87 ICBT sessions, 5.58 cognitive-training sessions, 2.87 approach-bias sessions, 6.35 reward redemptions and 16.17 check-ins. These numbers describe average recorded actions, not exposure doses or validated clinical improvements. Later app-use persistence cannot fill in missing forty-week craving, relapse or drug-concentration outcomes; those later biological and clinical endpoints were not collected in this study.

What happened at twenty-four and forty weeks?

At twenty-four weeks after enrollment, the reported overall attrition was 38.5%, meaning roughly three out of eight digital participants had completely ceased app use under the authors' definition. App usage did not necessarily stop across all modules at the same pace. The report described continued check-in, rewards and cognitive exercises among different proportions of users; remaining active in any one category is not identical to completing the prescribed eight-week therapy as designed. Its charts show that the more frequently used functions were logins and check-ins rather than intensive CBT modules.

By forty weeks, 44.2% of the app group was still active at least in the broad tracked sense, corresponding to 55.8% overall attrition. Average logged days approached 58 per user, whereas average completed ICBT lessons approached nine. These divergent activity measures suggest considerable variation in exposure to the package's components. Most importantly, the source does not furnish forty-week independently assessed urine toxicology, confirmed abstinence or behavioral-test efficacy. Using forty-week login data as proof of durable treatment benefit would cross the boundary from engagement into an outcome never measured.

Were app rewards, therapy, and assessment separable?

The intervention deliberately bundled instructional material, approach-training repetitions, cognitive games and points redeemable for cellular service. Because components were combined in a single treatment arm, the study could not independently determine which component was responsible for a favorable craving result. Nor could it establish how much improvement came from the convenience of private self-monitoring or from incentives for frequent activity. A future dismantling or factorial trial could separately test CBT, attention-bias contingencies, memory exercises, reminders and rewards while providing equivalent contact and assessment opportunities to each group.

This is particularly relevant when digital interventions are described with mechanistic vocabulary. Pushing a methamphetamine-associated image away on a screen is a motor response within an experimentally arranged stimulus contingency. The study did not measure dopamine-release concentrations in brain tissue, drug receptor occupancy, actual choice under intoxication, or the probability that a patient would refuse methamphetamine in the street. A claimed mechanism should be distinguished from what the researchers directly tested. Inference about abstinence must be anchored to abstinence measurements, not an assumed pathway from swiping behavior.

Can the findings be translated into a toxicological drug test?

No. The article provides no assay validation for estimating parent methamphetamine or amphetamine concentrations from cue-craving scores, app login counts, PHQ-9, GAD-7 or geometric recognition accuracy. There was no controlled methamphetamine administration by route or dose. Screening records were not used as an endpoint and were reported as uniformly negative during the relevant time period. Neither a negative screen nor a declining craving questionnaire can be algebraically converted into a measured concentration of drug in blood, a time-since-use estimate, or a statement about whether a named individual was intoxicated.

A genuine forensic toxicology analysis must identify the tested material and question. Venous whole blood contains cellular and plasma fractions and is not interchangeable with plasma or serum that has been separated by centrifugation or clotting. Urine and oral fluid provide different sampling windows and chemical contexts; hair, sweat and vitreous humor answer additional, nonidentical questions. Parent methamphetamine, amphetamine metabolite or co-exposure should be distinguished by a validated method and proper units if those compounds were actually measured. In this trial none of those participant-specific quantitative forensic specimens is documented. General interpretive principles are explained in the published same-site guides on biological matrices and confirmatory drug testing.

Could the trial prove safe driving, witness reliability, or a specific incident cause?

It could not. The participants did not perform an on-road driving assessment, validated drug-specific simulator challenge or eyewitness accuracy experiment. No identifiable person's behavior at a collision, workplace accident, arrest or other legal event was recorded. The interventions were psychosocial and digital, the outcomes primarily questionnaires and a brief visual recognition task, and the tested population already was entering structured community rehabilitation. A mean score difference in that group is not the same as the probability that an individual later operated a vehicle safely or truthfully reported an event.

Even if a study demonstrated a convincing randomized effect on craving, a separate evidentiary bridge would still be required to connect that population effect to a named person's actual exposure, contemporaneous symptoms, behavior, alternatives and relevant legal standard. Here the chain is shorter still: the source did not establish a group-by-time toxicology endpoint and did not directly assay drug concentrations. A technical report should avoid turning plausible digital-health benefits into unsupported individual conclusions, particularly where a claimed inference involves legal blame, criminal responsibility, safety-critical work or physical injury.

What are the principal study strengths and limitations?

A strength is that the investigators attempted randomized comparison in real community rehabilitation rather than presenting an uncontrolled user survey as proof of intervention effect. The article also reports several distinct outcomes and their genuine null findings, acknowledges attrition, and documents engagement rather than assuming everyone completed every component. Its paired and imputed analyses expose at least some uncertainty about missing data. These features make it useful as an early feasibility contribution and an example of how digital craving and memory measures may respond differently.

The limitations are substantial: one hundred participants but far fewer outcome completers, pronounced differential questionnaire response, unblinded participants, differing assessment settings and devices, a bundled intervention without matched sham engagement, and industry involvement in developing the application. Several original significance statements address within-arm paired change rather than a directly reported incremental randomized effect. Baseline and post-treatment urine drug-screening outcomes could not distinguish groups. The study enrolled a relatively selected, smartphone-capable community treatment population, not an acute emergency or driving cohort. These factors preclude a definitive claim that smartphone use reduced methamphetamine intake.

How should an attorney, scientist, or clinician describe the result?

A defensible summary is precise: a randomized community-treatment study assigned fifty-two people to an eight-week multi-component smartphone application and forty-eight to telephone-based usual care. Some app-arm self-reported cue-craving analyses showed statistically significant within-arm changes, and an imputed visual recognition estimate reached p = .041, whereas the complete-case recognition estimate was nonsignificant and depression and anxiety changes were also nonsignificant. Toxicology screening was not used to establish reduced consumption, and differential missing outcomes limit direct causal comparison.

The source is relevant to feasibility, engagement, missing-data methods and laboratory-measured cue-response outcomes. It cannot be cited as a blood concentration study, a validated relapse detector, proof of acute drug effects, or an estimate of a person's safe-driving capacity. For courts or expert reports, investigators should identify the original primary endpoint, method of analysis, study denominator and outcomes genuinely absent. Where patient-specific conclusions matter, relevant clinical records and properly validated specimen results must be examined separately; this publication does not substitute for them.

Frequently asked forensic and scientific questions

Did the investigators administer methamphetamine to participants?

No. The trial randomized a psychosocial smartphone program versus community support; no controlled methamphetamine dose, route or measured pharmacokinetic profile was administered.

Did the application prove that participants stopped using methamphetamine?

No. Urine/toxicology screens were not used as an efficacy endpoint because all 100 participants lacked a positive recorded screen in the relevant interval; the primary studied endpoint was picture-cue craving self-report.

How many independent participants were studied?

One hundred people were randomized: 52 to the digital application and 48 to treatment as usual. Repeated swipes, app sessions, or questionnaire items were not independent human participants.

Was digital treatment more effective than the control arm?

The paper reports significant within-app-group paired craving changes, but those are not themselves a formal statistically significant randomized group-by-time comparison. Unequal missing responses and assessment logistics restrict causal interpretation.

Did the app improve memory under both missing-data approaches?

No. The complete-case visual recognition result was nonsignificant (p = .094), while the imputed analysis reported an estimate of +0.220 and p = .041. The difference is method-sensitive.

Could the craving score identify methamphetamine in a blood sample?

No. It was a sum of self-ratings to thirty drug-related images, not an analytical chemistry method. The study reported no participant-specific validated blood methamphetamine concentration.

Can an expert use this study to determine driver impairment or collision causation?

No. It did not measure driving, collision circumstances, contemporaneous quantitative whole-blood drug, or person-specific exposure–response relationships.

Primary study and directly relevant published guides

Read the complete original empirical study: Zhang et al., 2022, DOI 10.3389/fpsyt.2022.1027695. For how measured chemical concentrations must remain associated with their specimen and analytical procedure, the separately published same-domain resources on biological matrices in forensic toxicology and confirmatory drug testing provide additional context. Neither guide converts this trial’s cue response score into an individual drug concentration.