Can Resting-State fMRI Predict Methamphetamine Craving After Treatment? The 82-Patient Study and Its 60% Accuracy Limit

Short answer: A 2026 primary study found that baseline resting-state functional connectivity among 82 adults with methamphetamine use disorder showed a statistically detectable but modest relationship with their subsequent self-reported reduction in craving during a residential treatment program. A nested cross-validated model produced a prediction-versus-observation correlation of r = 0.33 (permutation p = 0.0054). When the investigators converted the outcome into two categories, classification accuracy was 60%, with 59% sensitivity and 61% specificity. That is exploratory signal, not a validated individual prognosis. Crucially, the outcome was change on a craving questionnaire—not observed relapse, independently established continued drug abstinence, a forensic drug assay, or a person’s driving performance.

This analysis examines the actual original empirical paper by Du, Di, Luo and colleagues, Functional Brain Network Predictors of Abstinence Treatment Outcomes in Methamphetamine Use Disorder, published in CNS Neuroscience & Therapeutics (2026), primary study DOI. The defensible question is narrowly framed: what can an imaging algorithm predict about a one-year craving-score change in this cohort, and what does it leave unmeasured? It is not a comparison of MRI scanning against a legally collected, chemically analyzed drug specimen.

What did the investigators actually try to predict?

All 82 individuals with methamphetamine use disorder underwent baseline magnetic resonance imaging before a structured year-long program involving medical care, education about methamphetamine, and physical activity within drug rehabilitation centers. They completed a 25-item Methamphetamine Craving Questionnaire before and after the program. The authors expressed the outcome as (first score minus second score) divided by first score. Thus, a person who reports less craving obtains a more favorable percentage change. This continuous percentage, rather than a demonstrated laboratory-negative year or independently confirmed absence of relapse, was the prediction target.

The wording matters because a treatment center may describe its program as abstinence-based while the research endpoint remains a questionnaire. A measured change in urges, intention to use, positive expectations, or perceived self-control has psychological significance. It cannot by itself establish that a particular person did not consume methamphetamine during the interval. The article did not prospectively validate the brain model against toxicologically verified relapse events. Its authors expressly identified collecting actual relapse data as future research.

How many biologically independent participants were studied?

The main patient cohort contained 82 people, not the number of scans, MRI time points, imaging connections, cross-validation iterations, or genes. There were 68 separate healthy controls, bringing the overall participant count to 150. The prediction analysis used the 82 patients’ own baseline connectivity and later craving-score changes; the 68 controls were not 68 additional patients who had the same treatment outcome. Within the 82, an exploratory threshold divided 39 into a higher-craving-reduction group and 43 into the comparison group.

A 3-Tesla scanner acquired resting-state functional MRI using 225 image volumes and a brain parcellation of 246 regions. Those repeated images and brain regions are measurements nested within a human, not biological replications. For every outer leave-one-out validation iteration one patient was held aside; repeating that operation 82 times still interrogated the same initial 82-person cohort. This distinction is essential whenever large computational outputs make a modest human sample seem far larger.

Why is a 30% improvement threshold not the same as abstinence?

The investigators labeled patients with more than 30% reduction in the craving questionnaire as “responders,” and those with a reduction of 30% or less as “non-responders.” They specified that this binary threshold served secondary analyses; the primary prediction used continuous percentage change. Neither label denotes a chemically verified biological category. Someone can experience reduced cravings and still use a drug; someone can have persistent cravings without using one. The paper did not measure these two possibilities as alternative endpoints.

The percentage calculation also depends on the starting score. That matters because the groups already differed substantially at baseline. Responder-group mean craving was 67.03, compared with 56.02 among those below the 30% threshold, a reported p = 0.001. The arithmetic design makes baseline measurement error and regression toward a typical score scientifically relevant. It is therefore inappropriate to rebrand the dichotomy as a stand-alone relapse prediction test.

What do the original cohort and endpoint numbers show?

The first original evidence table places the actual independent headcounts and questionnaire measurements beside the major observed imbalance. Values are means with reported standard deviations where the source supplied them; they are not chemically determined drug levels. The significance values concern comparisons of the groups after researchers defined them by their questionnaire change.

Across all 82 patients, mean age was 34.43 years, 58 were men and 24 women; the program’s reported abstinence-period mean was 342.66 days. The separate healthy-control group contained 68 participants, of whom 46 were men, with mean age 40.10 years. Controls did not provide corresponding 12-month treatment-response labels. Counting controls as additional positive or negative examples would distort the reported design.

Why is the sex imbalance a major scientific concern?

The responder-defined group was approximately 95% male, whereas the other group was approximately 49% male. That difference is striking enough to affect any descriptive comparison of their imaging signals or psychological endpoints. It creates an alternative explanation that should be examined rather than suppressed: sex composition, or other variables correlated with sex in this selected sample, could contribute to apparent group-associated network signatures. Baseline craving scores also differed. The investigators reported additional group-versus-control analyses adjusted for gender, but an adjustment does not magically make the original groups exchangeable or prove that every model prediction was independent of demographic confounding.

It would be equally inaccurate to assert that this imbalance necessarily explains the entire imaging association. The proper statement is narrower: the sample has a highly uneven covariate distribution, so observed response-group contrasts require sensitivity analysis and replication in independent, adequately balanced cohorts. This uncertainty is unusually important for an article that might otherwise be quoted to suggest a future personalized forensic or clinical prediction tool.

What does resting-state functional connectivity measure?

Functional MRI detects time-varying blood-oxygen-level-dependent signals associated with local hemodynamics. Resting-state connectivity is a statistical relation between signal fluctuations in predefined regions while people are not performing a particular decision task. It does not directly sample neurotransmitter molecules, synaptic dopamine concentrations, or methamphetamine in the bloodstream. A correlation between two imaging-derived time series is not the same thing as an anatomical cable or a directly observed causal route between brain structures.

Here the investigators processed images with motion-related screening and other nuisance regression, temporal filtering, and standard-space registration. A 246-region atlas defined a large candidate space of connections. Principal component regression then reduced that feature space before mapping imaging features to the continuous craving-reduction score. A person may produce thousands of connection coefficients, yet the empirical prediction problem still depends on 82 patient outcomes. Modeling must address feature selection and correlated signals so apparent success is not simply a function of too many candidate predictors.

How did nested validation and permutation testing work?

For the outer validation loop, the model left one patient out, trained on the other patients, and predicted the held-out score; it repeated the procedure until every patient had a held-out prediction. Within training data, an inner procedure selected the principal-component setting. That nesting is more defensible than tuning model complexity on a person’s test observation. Across iterations, retained components accounted for a reported mean 87.35% of variance in the imaging features, with a 2.61 percentage-point variation. That feature-variance percentage is not 87% of eventual craving outcomes explained.

For significance assessment, the investigators broke the correspondence between baseline imaging and outcome labels 5,000 times, rebuilt the prediction procedure under those shuffled correspondences, and compared the observed correlation to the resulting null distribution. The permutation result of p = 0.0054 provides evidence that the model’s prediction association was greater than expected under that test’s exchangeability assumptions. It does not prove that a future hospital, another scanner, a new ethnicity or treatment setting, or an individual defendant would obtain comparable accuracy. The paper itself states that its model had not been tested on new methamphetamine-treatment subjects.

How strong was the continuous prediction result?

The main out-of-fold prediction-versus-actual outcome correlation was r = 0.33. This is a positive but moderate association. The statistic does not mean that the algorithm correctly predicted 33% of participants, reduced cravings by 33%, or identified drug use with 33% sensitivity. Nor should its numerical square be presented uncritically as an externally validated estimate of actionable individual prognosis. It quantifies association in the evaluated sample under the specified pipeline.

The source reported an approximately zero mean residual (0.0014) and a Breusch–Pagan test p = 0.22. That latter result was nonsignificant for detected heteroscedasticity; it is not a guarantee that every model assumption is true. A residual diagnostic and a permutation p-value answer different questions from whether errors are sufficiently small for real-world decisions. The study did not establish clinically acceptable predictive error thresholds or prespecified consequences of mistaken responder assignments.

What happened when the investigators made a yes-or-no prediction?

For secondary analysis, the 30%-craving-change categories were compared with a rule based on model-derived response-network strength. The healthy-control mean served as a threshold when distinguishing the two patient classes. The report gives accuracy 0.60, sensitivity 0.59, and specificity 0.61 in a repeated internal cross-validation procedure. These figures have a familiar practical meaning: correctly identifying around six cases out of ten in the specific modeling framework leaves substantial errors in either direction. The metrics cannot serve as a stand-alone forensic or medical classification rule.

An approximately balanced binary sample with 39 responders and 43 nonresponders makes 60% only a modest improvement over choosing the larger category for every patient. Sensitivity and specificity concern questionnaire-defined labels, not truth about post-treatment drug use. Before using an algorithm for consequential decisions, an independent sample should test calibration, decision thresholds, sex and age performance, scanner transportability, and actual validated outcomes. Those results are not available in this primary study.

Which key prediction and genuine null findings belong side by side?

The second original evidence table separates measured significant associations from the source’s actual tested nonsignificant relationships and the distinct question of external validity. Its numbers should not be combined into a single generalized accuracy claim. Statistical significance for one outcome does not convert every neighboring null clinical variable into proof of predictive usefulness.

Age at first methamphetamine use also had a tested nonsignificant association with the craving response score (r = 0.17; p = 0.13). These null results are not evidence that age, previous drug exposure, nicotine, or alcohol could never matter in a larger or differently measured population. They show that this sample did not establish those specified simple relationships. Importantly, a descriptive failure to reach a threshold is not an equivalence or noninferiority test.

Which brain networks contributed to the model?

The strongest positively weighted connections involved visual-network coherence, frontoparietal-to-default-mode relations, and visual-to-dorsal-attention relations. Negatively weighted contributions included somatomotor within-network connections, somatomotor-to-default-mode links, and default-mode-to-ventral-attention links. At a regional level, weights implicated portions of frontal cortex, orbital cortex, parahippocampal structures, visual cortex, and other regions. These are coefficients contributing to a fitted prediction; their signs are not chemically measured increases and decreases in dopamine.

It is particularly risky to describe the entire default mode network or orbitofrontal cortex as uniformly “good” or “bad.” A given region participates in multiple edges with different signs. The original authors observed that orbital areas appeared in positive and negative predictive patterns, consistent with circuit heterogeneity rather than a one-dimensional dial of addiction severity. A brain network’s contribution to predicting a score does not establish which particular neural connection caused that person’s future behavior.

Why did the healthy controls show an apparently paradoxical result?

When all methamphetamine participants were compared with the 68 controls together, the global response-network strength difference was nonsignificant (t = −0.52, p = 0.60). When the patient sample was split according to its later craving-score reduction, the higher-response group showed greater modeled network strength than controls (t = 6.03, p = 2.5 × 10⁻⁸), whereas the other group showed lower network strength (t = −5.56, p = 1.9 × 10⁻⁷). Group contrasts remained significant in the reported sex-adjusted analyses, with p = 0.0059 and p = 0.0017 respectively.

This juxtaposition is informative: combining differently selected patients can conceal differences seen after stratification. But the categories depend on a later self-report threshold, so the contrasting patterns should not be mistaken for diagnostic brain states objectively observed before a relapse or an abstinence event. The source control sample did not independently validate the longitudinal treatment-response algorithm. The entire cohort and the conditional subgroups answer different statistical questions.

Are stable model weights the same as external validation?

The investigators reported model-weight correlations of about 0.94 ± 0.04 across leave-one-out iterations, 0.85 for a bootstrapped comparison, and 0.87 for comparison with partial least squares regression. These checks speak to internal similarity of fitted coefficient patterns when resampling or changing the modeling approach. They are useful engineering diagnostics but do not demonstrate reliable out-of-hospital predictions in previously unseen clinical populations.

Likewise, removing the highest-weighted network edges lowered in-sample predictive performance, which is consistent with those features carrying information for the fitted model. It does not prove a physiological manipulation of those connections would change future craving. The distinction between measurement, association, statistical stability, and intervention is central both to scientific reporting and to testimony that must be limited to what was actually tested.

What were the gene-expression findings—and whose tissue was analyzed?

To interpret the imaging-derived network, the study matched regional imaging patterns against a separate Allen Human Brain Atlas transcriptomic reference. The atlas information derives from six donor brains and thousands of sampled tissue regions, not from 3,702 additional patients with methamphetamine use disorder. Gene lists of 780 positive and 780 negative associations were constructed from ranked correlations, with false-discovery-rate procedures reported for those regional correlations. The positive list showed enrichment for neurotransmission-related pathways and inhibitory-neuron markers, while the negative list featured other cellular patterns.

Such enrichment is a hypothesis-generating map from a public reference dataset. It is not direct evidence that the 82 patients had measured dopamine deficits, a validated biological relapse signature, altered transmitter concentrations, or a genetically determined treatment outcome. The imaging patients did not undergo matching brain biopsies. Six postmortem donor brains do not transform an observational clinical connectivity study into an individual molecular assay.

What biological specimens were actually tested for methamphetamine?

The inclusion process used a categorical urine drug screen: the patient group was required to have a positive result for methamphetamine and negative results for other screened substances, while controls screened negative. That is a selection criterion, not a time-resolved quantitative methamphetamine exposure measurement for a legal case. The article did not report validated numerical methamphetamine or amphetamine concentrations from human venous whole blood, separated serum or plasma, oral fluid, hair, sweat, vitreous, milk, breath condensate, or forensic tissue. Neither an imaging signal nor a self-report scale supplies missing chromatography.

Forensic interpretation first needs the precise collection matrix. Human venous whole blood is not interchangeable with centrifuged plasma or serum: analyte distribution, analytical units, sample handling, and validated concentration-to-effect claims can differ. A urine screening category likewise does not give an ng/mL whole-blood concentration or exact time of consumption. For a broader orientation, the site has separate published guides to differences among toxicology specimen matrices and to confirmatory analytical drug testing. Those general guides are context; neither adds unmeasured results to this experiment.

Was a drug dose administered or chemically confirmed in this study?

No experimental methamphetamine dose was administered to establish a concentration–response curve. Participants’ reported prior amounts of street methamphetamine in grams per day were historical use estimates, not controlled doses of a specified chemical form with known purity, salt identity, or route. Those reports also lacked a validated analytical link to parent methamphetamine or amphetamine metabolite measurements in a contemporaneous biological specimen. Historical reported use and a future craving-score change cannot define someone’s concentration at a particular day or event.

This distinction also rules out transferring findings from controlled oral d-amphetamine or prescribed lisdexamfetamine trials to methamphetamine casework. They involve different experimental compounds, administrations, cohorts, analytical endpoints, and sometimes entirely different biological matrices. A predictive imaging paper does not become drug pharmacokinetics simply because its sample has a history of stimulant use.

Could the model establish relapse, safe driving, or witness performance?

It cannot on the evidence reported. Actual subsequent relapse incidence was not collected as the model’s validated outcome, and laboratory-confirmed long-term nonuse was not the primary endpoint. No driving simulator, observed road test, collision investigation, contemporaneous structured sobriety examination, witness-memory task, or legal event-specific causation analysis supplied a target for the prediction. Even an imaging model that had stronger clinical performance would require separate validation before relating it to those new contexts.

Two independent limitations apply in an individual case. First, the study’s measured phenomenon is different from the proposition usually at issue: change in craving versus actual intoxication or impairment. Second, a modest population-level predictive correlation does not yield a defensible probability that a particular person had consumed methamphetamine at a specified time. An attorney or scientist should preserve the difference between an interesting group association and an admissible, case-specific analytical inference.

What do nonrandom treatment, selection, and generalizability mean here?

This was an observational prediction study within real rehabilitation programs, not a randomized test assigning people to different methamphetamine exposures or experimentally changing network connectivity. Everyone in the patient cohort participated in the center’s treatment setting. The prediction therefore cannot separate the contribution of imaging predictors from features of treatment access, selection for participation, withdrawal history, or unmeasured individual factors. Screening for psychiatric exclusions relied partly on interview rather than a complete standardized scale battery, an acknowledged limitation.

The sex difference is not a small footnote. Nor is the moderate sample size, the same-site cross-validation, or the potential sensitivity of functional MRI to acquisition and preprocessing changes. The patient group was younger on average than the healthy controls, and histories of nicotine and alcohol varied. These circumstances do not negate a real observed correlation; they narrow the reliable scope of any extension to a new patient, a different scanner, a new treatment environment, or a forensic population with other exposure and illness histories.

How should experts describe the study without overclaiming?

A fair summary states the actual original outcome, the biologically correct sample size, and both positive and null evidence. Baseline imaging from 82 treated adults correlated modestly with a later questionnaire-derived craving change under a nested internal validation method. Secondary categorical classification was 60% accurate, and a dramatic sex imbalance existed between the defined outcome groups. None of that establishes a drug-positive whole-blood sample, observed relapse, individual legal intoxication, or crash causation. The source’s mechanistic suggestions about reward circuitry and neurotransmitters are interpretations, not directly assayed clinical facts.

The best scientific next test would be independent, multisite, prospectively preregistered validation with balanced participant groups, transparent failure rates, quantitative external calibration, and outcomes separating craving ratings from objectively monitored drug-use or relapse events. If an investigation involves drug concentration, it additionally requires specimen-specific validated analytical toxicology. Those future experiments would extend rather than retroactively change what this paper measured.

Seven evidence-based questions about the 82-patient fMRI study

1. Did the MRI algorithm prove that a patient stayed abstinent for one year?

No. The algorithm predicted a percentage change on a 25-item methamphetamine-craving questionnaire after a year-long residential program. It did not use independently confirmed relapse or serial quantitative toxicology as its validated outcome. Calling a questionnaire-derived improvement “verified abstinence” would misidentify the measured endpoint.

2. How many participants contributed to the craving prediction?

Eighty-two patients contributed baseline imaging and later craving outcomes to the central model. Sixty-eight healthy controls were a distinct comparison cohort; MRI frames, cross-validation folds, atlas regions, and donor tissue samples were not additional patient observations. The analyzed patient denominator remains 82.

3. Was a statistically significant correlation of 0.33 clinically accurate enough?

Statistical significance and consequential classification accuracy are different. The correlation r = 0.33 had permutation p = 0.0054 under the authors’ internal test, while the secondary group classifier had accuracy 0.60, sensitivity 0.59, and specificity 0.61. Independent clinical and forensic validation was not established.

4. Why does the 37-to-2 versus 21-to-22 sex split matter?

Participants defined as responders were 37 men and two women, compared with 21 men and 22 women in the other group, p < 0.001. That difference threatens straightforward interpretation of contrasts between the two outcome groups. Reported gender-adjusted comparisons do not eliminate the need for independent balanced-sample testing.

5. Did investigators measure methamphetamine concentration in venous whole blood?

No. A categorical urine drug screen was described as an eligibility criterion. The study did not report quantitative parent-drug or metabolite assays of legally collected human whole blood or of separated plasma or serum. fMRI connectivity and craving scores are not substitutes for those chemical measurements.

6. Did baseline drug-use history predict the questionnaire improvement?

The tested simple clinical correlations were nonsignificant: prior duration of use p = 0.25, age at first use p = 0.13, reported grams/day p = 0.31, nicotine-dependence score p = 0.70, and alcohol-use score p = 0.28. A nonsignificant result does not prove that a factor has no possible effect in other populations.

7. Can this paper support a conclusion about an individual driver’s impairment or a collision?

No. It contained no acute controlled methamphetamine administration, validated whole-blood chemical concentration, observed driving task, collision reconstruction, or case-specific event measurement. Even a statistically supported population-level prediction of self-reported craving cannot identify an individual’s drug exposure, impairment, or causal contribution to an incident.

Bottom line: a measurable group signal, not a forensic drug test

The genuine 2026 empirical finding is an internally cross-validated association between baseline resting-state connectivity and later questionnaire-defined craving reduction in 82 patients, with a moderate r = 0.33 and modest 60% secondary classification accuracy. The study’s strong sex imbalance, nonsignificant clinical predictors, and lack of external validation are indispensable context, not details to omit for a more striking headline. No drug concentration, confirmed longitudinal relapse status, individual impairment, or event-specific causal conclusion was measured. The original published primary research record should be consulted for the defined endpoints and study design rather than treated as a general-purpose forensic prediction tool.