What Are You Craving? Using Wearables to Distinguish Food and Drug Cravings During Treatment with Extended-Release Buprenorphine

Abstract

Craving, or the subjective, strong desire to use a substance, is a central factor in addiction, and part of the diagnostic criteria for substance use disorders (SUDs). Cravings can also occur for other triggers such as food, and cravings for food and drugs have been found to activate distinct neural pathways in the brain. Recently, physiologic signals from wearable devices have been applied to digitally detect cravings in patients with SUDs. But to date, no studies have explored digital detection of cravings by subtype. We collected continuous physiologic sensor data from N = 12 participants with opioid use disorder (OUD), treated with extended-release buprenorphine (BUP-XR). Data were analyzed to assess whether sensor signals carried differential information that could distinguish between food-, drug- and mixed-craving types. Accelerometer, heart rate and heart rate variability features significantly differed between drug, food and mixed trigger cravings. Cross validated models trained with these features distinguished each type of craving with area under ROC curve ranging from 75%–80%. These findings support the ability of wearable sensor-based digital biomarkers to distinguish craving subtypes in individuals with OUD.

1. Introduction

In 2023, drug overdoses accounted for an estimated 105,000 deaths in the United States (Garnett et al., 2024), making it the leading cause of accidental death in the country (Curtin, 2025). While these deaths may only represent a fraction of the deaths from the most significant contributors to mortality in the USA, which are heart disease and cancer (702,880 and 608,371, respectively) (Kochanek et al., 2023), they represent an entirely preventable cause of death that disproportionately affects younger individuals (Garnett et al., 2024). Efforts to improve our understanding of substance use disorder (SUD) to better treat and thereby reduce the mortality of those affected are therefore an essential focus of research to address this public health crisis.

The concept of craving is vital to our understanding of SUDs. Definitions of craving vary by source, but it is generally described as a subjective, strong desire to use a substance. This implies that craving is conscious, represents an intense want, and is specific to an individual’s substance of choice (Tiffany & Wray, 2012). Craving is considered a central factor in addiction, aligning with its inclusion in diagnostic criteria for most SUDs (Association, 2013). A substantial body of literature supports craving as significantly associated with both the diagnosis and outcomes of various SUDs (Kakko et al., 2019; M A Sayette et al., 2000; Tapper, 2018; Tiffany & Wray, 2012). Cravings are also recognized as one of the strongest predictors of return to drug use during treatment (Kavanagh & Connor, 2013; Preston et al., 2018; M A Sayette et al., 2000; Serre et al., 2015; Witkiewitz et al., 2013). Consequently, cravings have become targets for therapeutic intervention and are important outcomes in related clinical trials (Michael A Sayette, 2016; Tiffany & Wray, 2012; U.S. Food and Drug Administration, 2020). A deeper understanding of cravings, including how they respond to medications and other interventions, is critical for improving and personalizing SUD treatment.

Despite validated tools existing, there is no universally accepted questionnaire or rating scale for measuring cravings (Tiffany & Wray, 2012). Craving experiences are inherently subjective, requiring patients to recognize, acknowledge, and assess their cravings’ intensity. Recent research has explored objective craving detection through wearable sensors (Carreiro et al., 2020, 2021, 2024; Shrestha et al., 2023). Initial studies using research-grade sensors (Empatica E4, Empatica, Milan, Italy) demonstrated a 75.7% accuracy in distinguishing self-reported craving from non-craving states using physiologic data (e.g. heart rate, accelerometry parameters, skin conductance) (Carreiro et al., 2020). Although these sensors were costly and bulky compared to consumer products, subsequent research showed that applying similar algorithms to data from consumer-grade wearables (Garmin Vivosmart^® 4, Garmin Ltd, Olathe, KS) achieved 71% accuracy (Shrestha et al., 2023). Concurrently, the “Realize, Analyze, Engage (RAE)” digital health app was developed to integrate consumer-grade wearable sensor data into a therapeutic tool to enhance SUD recovery (Carreiro et al., 2021). This app detects stress and cravings, triggering dialectical behavioral therapy (DBT) interventions to promote mindfulness and aggregates data into a clinical portal accessible by clinicians and patients (Carreiro et al., 2021). Recent studies indicate that the RAE app maintains a similar accuracy level in predicting cravings and engagement with the app is associated with lower rates of return to drug use (Carreiro et al., 2024).

Cravings are not exclusive to SUDs but extend across various domains, including food, gambling, video games, social media, the internet, television, and sexual activity. Behaviorally, these cravings can be powerfully cue-induced, leading to compulsive behaviors and overconsumption triggered by environmental stimuli like sights, smells, or paraphernalia (Rogers, 2017). Enhancing our understanding of craving mechanisms can thus help individuals manage unhealthy or undesired cravings across these domains. Additionally, better distinguishing drug from other craving triggers could help improve the accuracy of wearable-based detection of craving from the current 69%–71% accuracy obtained in the studies discussed above (Carreiro et al., 2024; Shrestha et al., 2023). Food cravings, in particular, are significant due to their frequency and the literature suggesting similar neural activations as drug cravings (Kronberg & Goldstein, 2023). Evidence indicates that individuals might confuse food cravings with alcohol cravings, potentially undermining recovery efforts (Czarnecki et al., 2023). This confusion likely applies to other substances of misuse as well. If wearable sensors can assist in distinguishing between drug-related and other cravings or identify drug cravings with improved accuracy, digital tools like the RAE app (Carreiro et al., 2021) could provide targeted prompts, potentially directing individuals toward fulfilling non-substance-related cravings (e.g., food) rather than returning to drug use. In the present study, we tested the hypothesis that sensor data features carried differential information about craving types in a population with a common SUD, opioid use disorder (OUD). We compared physiologic data from a wrist worn wearable sensor during drug and non-drug cravings to 1) evaluate differences in sensor features, and 2) investigate the feasibility of leveraging machine learning models to distinguish between the types of cravings, reported by users as drug-related, food-related and mixed cravings.

2. Methods

This is a secondary analysis of digitally detected craving events collected from individuals with OUD receiving treatment extended-release buprenorphine (BUP-XR), a once-monthly injectable partial opioid-agonist. The parent study and all related activities were reviewed and approved by the Institutional Review Board (IRB) at the senior author’s institution.

2.1. General Study Protocol

Study participants were recruited from local primary care clinics (via word of mouth and electronic flyers) and through an online research portal (ResearchMatch.org). Individuals were considered eligible if they were: adults (18 years of age or older), able to read English, able to provide informed consent, enrolled in a treatment program for OUD and either had a plan to initiate BUP-XR or had started the treatment within past six months, and had access to a smartphone capable of running the RAE Health app. Individuals were excluded if they: had limitation of range of motion of their non-dominant arm (e.g. fracture, amputation), were considered prisoners, or were pregnant.

Once informed consent was obtained, participants were given a brief device training session on the wearable device (Garmin Vivosmart^® 4) and the RAE Health App (ContinueYou, LLC, Bristol, ME). Participants were then asked to use the wearable device and app continuously for a period of 60 days, and only to remove the sensor when charging. They were asked to keep the RAE app running in the background on their phone at all times, and to respond to all app prompts.

The analysis consisted of multiple steps including categorizing events into distinct types of cravings based on the participant responses entered for each craving, estimating sensor features from signals recorded by the wearable device when a craving was detected by the app, and developing models to discriminate among types of cravings.

2.2. Wearable device and mobile app

The Garmin Vivosmart^® 4 (Figure 1) is a low-profile, consumer-grade smartwatch. The device collects physiologic data through its noninvasive sensors (heart rate monitor, barometric altimeter, accelerometer, and pulse oximeter) and produces (through the Garmin application programming interface) additional calculated metrics for physical activity and sleep. The Vivosmart^® 4 syncs data to the RAE Health app when in Bluetooth range (30 feet or 10 meters).

Figure 1. Garmin.

The iOS/Android compatible RAE Health app (Figure 2) continuously received raw sensor data from the Garmin sensor, which was then extracted by the investigators. Embedded in the app were previously validated machine learning algorithms which detect craving (Carreiro, et al., 2020). The app algorithms were derived and validated on people with SUD and with mixed drug craving (including opioids, stimulants, sedatives and alcohol). When a craving event was detected, a push notification was delivered to the user’s phone, asking them to confirm or deny if a craving has occurred, rate the craving on a visual analog scale of 1–5, and provide context for the craving (including what they are craving). Standard options included “Alcohol”, “Caffeine”, “Food”, “Nicotine”, “Sugar” and “Drugs” (Figure 2). If “Drugs” was selected, the user was then asked to specify the drug type. App users were allowed to select greater than one craving trigger (e.g. “check all that apply”). App users also had the opportunity to manually initiate a craving event in the scenario that they perceive a craving that the algorithm did not detect. Study participants were encouraged to manually enter cravings if they experienced them (and did not receive an automatic notification). Once initiated, manually detected craving events provided the same rating and category prompts as automatically detected cravings. Instructions were provided at the start of the study, and reminders were provided at three study check in points (study day 20, 30 and 40).

Figure 2. RAE Health App Screenshots.

2.3. Grouping cravings by type

Annotations for craving alerts were mostly single words, such as “food”, “sugar” or “cocaine”. If participants selected more than one trigger, then a craving event was annotated with multiple words. Examples include: “nicotine, food, sugar” or “caffeine, nicotine, fentanyl, sugar”. A set of terms describing the drug-related and food-related categories were compiled and compared with each response, to assign a craving event to the correct category.

Craving events with user responses were split into three groups related to only food (FC), only drug (DC) and mixed cravings (MC). These groups were formed by comparing each response to a set of terms describing the categories. If a response had any of the terms included in the drug-related or DC set, it was marked as a ‘DRUG’ craving. Similarly, a response matching any term in the food terms set, was noted as a ‘FOOD’ craving or FC. Users sometimes provided several words to describe a craving, as given in the examples above, with both food or/and drug related terms along with “NICOTINE”. This ambiguity in the underlying cause/reason for the craving was accounted for by grouping these as ‘MIXED’ cravings, MC. Because nicotine could be considered as a drug for its addictive nature, cravings to which the user responded with both nicotine and one or more terms from the drugs set, were excluded from the analysis, for the sake of simplicity. Hence, the MC category consisted of nicotine with one or more food related terms.

2.4. Sensor data analysis

The data corresponding to the accelerometer, heart rate and inter (heart) beat interval recorded by the device over a five-minute period during which a craving was detected were considered. To be included as an instance in the analysis, each craving event needed to have all three data types recorded for five minutes including the time instance when the alert was generated and the corresponding notification was sent to the subject by the RAE Health app. In some cases, there was data loss when the device failed to transfer the data (either all three types or a subset) to the data server after event detection. Such a lack of sensor data would lead to the exclusion of a detected event from further analysis.

Features in the temporal, spectral and nonlinear domains were estimated from the sensor data (Shaffer & Ginsberg, 2017; Zhu et al., 2017). The estimated features are listed in Tables 1 and 2. These features were compared pairwise between craving groups - DC, FC and MC by non-parametric rank sum test, suited for small samples and significance identified by p<0.05.

Table 1.

List of accelerometer (ACC) and heart rate (HR) features

	Features		Abbreviation
Time domain features	Mean		mean_i, i =X, Y, Z for ACC mean, for HR
	Standard deviation		std_i, i =X, Y, Z for ACC Mean, for HR
	Skewness		skew_i, i =X, Y, Z for ACC skew, for HR
	Kurtosis		kurt_i, i =X, Y, Z for ACC kurt, for HR
	Minimum		min_i, i =X, Y, Z for ACC min, for HR
	Maximum		max_i, i =X, Y, Z for ACC max, for HR
	Energy		Energy_i, i =X, Y, Z for ACC Energy, for HR
	Root mean square (rms)		rms_i, i =X, Y, Z for ACC rms, for HR
	Mean absolute error		mae_i, i =X, Y, Z for ACC mae, for HR
Spectral features	Mean spectral power density		mnP_i, i =X, Y, Z for Acc.
	Std. deviation of spectral power density		sdP_i, i =X, Y, Z for Acc.
	Median spectral power density		mdP_i, i =X, Y, Z for Acc.
	Area under the power density curve		arP_i, i =X, Y, Z for Acc.
Nonlinear features	Sample entropy	Dimension =1	SmEn1_i, i =X, Y, Z for Acc. SmEn1, for HR
		Dimension =2	SmEn2_i, i =X, Y, Z for Acc. SmEn2, for HR

Table 2.

List of heart rate variability features

			Abbreviation
Time domain features	Standard deviation of NN intervals		SDNN
	Standard deviation of successive interval differences		SDSD
	Root mean square of successive interval differences		RMSSD
	Proportion of intervals that differ by more than x msec	X= 50 msec	pNN50
		X=100 msec	pNN100
	Integral of the density of the RR interval histogram divided by its height		TRI
	Baseline width of the RR interval histogram		TINN
Spectral features	Relative power of the low-frequency band (0.04–0.15 Hz)		PLF
	Relative power of the high-frequency band (0.15–0.4 Hz)		pHF
	Ratio of LF-to-HF power		LFHFratio
Non-linear Features	Approximate entropy		ApEn
	Poincaré plot standard deviation perpendicular the line of identity		SD1
	Poincaré plot standard deviation along the line of identity		SD2
	Ratio of SD1-to-SD2		ratio

With the limited number of cravings in each set, instead of a multiclass model, three binary classifiers which attempted to classify each group when the other two formed the alternate class were considered. Specifically, classifiers to distinguish DC when FC and MC were grouped as the other (majority) class, FC vs (DC and MC) and finally, MC vs (DC and FC) were developed. These models were cross validated, the receiver operating characteristics (ROC) curve was plotted, and the area under the curve (AUC) was used as a metric of the distinguishing ability of the model between the classes.

The features considered for machine learning were those with p<0.05, when compared between the two classes (e.g.: DC vs. (FC and MC) grouped). Principal component analysis (PCA) was carried out to reduce dimensionality and address multicollinearity (Gárate-Escamila et al., 2020). Components which together explained 95% of the total variance were retained and used to train machine learning classifiers, specifically, logistic regression (Log. Reg.), k nearest neighbors (kNN), support vector classifiers with linear (SVM_L) and non-linear (radial basis function) (SVM_RBF) kernels. These were cross validated, and their performance assessed by evaluating metrics of accuracy and AUC.

3. Results

3.1. Study participant information

A total of N=12 participants were enrolled, who continuously wore the wearable sensor device for 60 days. Table 3 displays characteristics of the study participants. A total of 153 craving events were recorded. Data loss occurred either when the device was not worn properly or lost during transfer from device to storage server. This caused one or more of the accelerometer, heart rate or inter beat interval data to not be available for analysis and hence, exclusion of the corresponding events. After excluding such events with missing data, 103 cravings had the full set of sensor data available for analysis.

Table 3.

Participant demographic information

Characteristic	N=12
Age (mean, range)	36.5 (18–57)
Sex (N, %)
Male	4 (33.3)
Female	8 (66.7)
Race (N, %)
White	11 (91.7)
Asian	1 (8.3)
Ethnicity (N, %)
Latino/a	2 (6.7)
Monk Skin Tone (N, %)
1	2(16.7)
2	1 (8.3)
3	6 (50)
4	2 (16.7)
5	1 (8,3)
Co-Occurring Psychiatric Diagnosis (N, %)	9 (75)
In treatment for polysubstance use (N, %)	8(68)

As described in the methods section, the responses to craving alerts were used to categorize each as belonging to a particular type, by comparison with a set of terms signifying each of the three types-drug (DC), food (FC) or mixed (MC) craving. There were 29 DC, 30 FC and 40 MC. Four MC were excluded for having multiple terms from both the food and drug category listed along with ‘nicotine’ in the user response.

3.2. Comparison of sensor features

Sensor features were estimated for each craving event for the three types of data collected by the wearable device – accelerometer, heart rate and inter beat intervals. These features, described in Table 1 and Table 2, were compared pairwise between the craving groups and are tabulated in Table 4 and Table 5. DC had consistently higher values for the significant accelerometer features than FC and MC. These were the only features that differed between DC and FC, whereas all heart rate and HRV features were comparable between DC and FC. Interestingly, DC and MC accelerometer features were also comparable.

Table 4.

Accelerometer features (Mean (SD)) compared between cravings by type

Feature	Drug Cravings (DC) N=29	Food Cravings (FC) N=30	Mixed Cravings (MC) N=40
SmEn1_X	1.92b, c (0.46)	1.59 (0.49)	1.47 (0.60)
SmEn2_X	0.67b, c (0.55)	0.33 (0.38)	0.35 (0.53)
mdP_Y	4084b, c (13431)	492.1 (682.3)	637 (1003.1)
SmEn1_Y	1.95b, c (0.52)	1.72(0.48)	1.59(0.56)
SmEn2_Y	0.70b, c (0.56)	0.39(0.52)	0.36(0.46)
mean_Z	−225.9b, c (696.0)	−658.8 (295.1)	−627.1 (315.7)
skew_Z	0.21b, c (2.83)	1.79(2.56)	1.57(2.12)
max_Z	772.87c (567.34)	484.18 (988.2)	444.9 (732.0)
mdP_Z	3.0e3b (1.3e4)	456.8 (811.1)	540.6 (1254.7)
arP_Z	9.7e4 (3.1e4)	1.05e5d (3.2e4)	8.9e4 (3.2e4)

^bp<0.05 between DC and FC

^cp<0.05 between DC and MC

^dp<0.05 between FC and MC.

Table 5.

Heart rate and HRV features (Mean (SD)) compared between cravings by type

Feature	Drug Cravings (DC) N=29	Food Cravings (FC) N=30	Mixed Cravings (MC) N=40
Heart rate features
mean	90.02 c (16.13)	83.30 d (13.99)	68.74 (14.85)
min	81.07 c (15.62)	74.50 d (13.56)	61.00 (14.19)
max	100.52 c (15.91)	94.23 d (16.09)	79.53 (16.20)
energy	1.9e6c (6.3e5)	1.6e6d (7.5e5)	1.2e6 (9.0e5)
rms	90.16 c (16.14)	83.50 d (14.00)	68.85 (14.84)
SmEn1	2.37 (0.38)	2.11 d (0.57)	2.34 (0.37)
SmEn2	1.29 c (0.34)	1.27 d (0.39)	1.58 (0.20)
HRV features
sdsd	0.03c (0.02)	0.03 d (0.02)	0.06 (0.03)
sdnn	0.05 c (0.02)	0.05 d (0.04)	0.07 (0.03)
rmssd	0.03 c (0.02)	0.03 d (0.02)	0.06 (0.03)
pNN50	0.08 c (0.12)	0.10 d (0.11)	0.24 (0.18)
pNN100	0.03 c (0.07)	0.03 d (0.05)	0.10 (0.11)
TRI	9.03 c (3.23)	10.11 d (6.96)	13.15 (4.97)
TINN	0.13 c (0.04)	0.14 d (0.10)	0.18 (0.09)
SD1	0.02 c (0.02)	0.02 d (0.01)	0.04 (0.02)
SD2	0.06 c (0.03)	0.07 d (0.06)	0.09 (0.05)
ratio	0.35 c (0.10)	0.44 (0.27)	0.45 (0.15)

^bp<0.05 between DC and FC

^cp<0.05 between DC and MC

^dp<0.05 between FC and MC.

Heart rate features of MC were significantly lower than both DC and FC, while HRV values were higher in the MC category, evidence of distinctive fluctuations in cardiac dynamics during these types of cravings. Of note was the observation that MC formed a unique craving type, with several features distinguishing it from DC and FC groups, despite having tags that could place it in either group. However, although the accelerometer features of MC differed from DC, they were comparable to FC.

3.3. Craving group classification by machine learning

Binary classifiers for each of the three craving groups as the minority class were developed and assessed for performance, using significant sensor-based features. The classifiers were developed with default parameters and assessed by 8-fold cross validation.

In the case of DC versus the majority grouping of FC and MC, 24 features, mostly in the time domain of the three sensor data types, differed significantly between the classes. After PCA, 10 components were found to explain 95% of the variance and were used to develop the classifiers. The ROC curves for the four classifiers with their optimal operating points are shown in Figure 3.

Figure 3. ROC curves of the classifiers to detect Drug Craving (DC) vs. grouped Food Craving (FC) and Mixed Craving (MC) class.

The models had comparable discriminative ability between the groups with the highest area under ROC of 0.76 for the SVM_RBF and Log.Reg. having the least with 0.72. The accuracy of classification ranged from 80% for kNN to 76% for the two SVM models. While specificity was higher than 90% for all the models, at the optimal operating points, the highest sensitivity for DC class was 0.62 for SVM_L and 0.31 for kNN.

When FC was considered for classification against the combined DC and MC groups forming the majority class, 12 features differed significantly between the classes. After PCA, six components were used to develop the classifier models of which kNN exhibited the best performance with area under ROC of 0.73 and accuracy of 75%. Log. Reg. had close discriminative ability at 0.70 AUROC and 70% accuracy. SVM classifiers had slightly lower area under ROC of 0.65 and 0.66 (Figure 4). Though the models had high specificity, the highest sensitivity of FC classification was 47% at the optimal operating point of the Log. Reg. model.

Figure 4. ROC curves of the classifiers to detect Food Craving (FC) vs. grouped Drug Craving (DC) and Mixed Craving (MC) class.

Hence though the models performed better than random classifiers, the classification of FC events was less successful than that of DC.

Finally, ML classifiers were developed to distinguish MC from the alternative class formed by grouping DC and FC observations. With 40 MC and 59 alternative class instances, this data set was the most balanced among the three considered for modeling. Twenty-one sensor features were obtained with p<0.05 when compared between the classes and following PCA, eight components were included in model development. Cross-validated models exhibited moderately high AUC ranging from 0.81 for the SVM_L model, 0.78 and 0.76 for the Log. Reg. and SVM_RBF respectively, to 0.73 for the kNN, as shown in Figure 5. The sensitivity and specificity of classification were also more balanced in this case, with the kNN exhibiting 73% sensitivity with 60% specificity and SVM_L providing 68% sensitivity for 83% specificity.

Figure 5. ROC curves of the classifiers to detect Mixed Craving (MC) vs. grouped Drug Craving (DC) and Food Craving (FC) class.

The classification of cravings by type, identified mixed cravings as a unique category that exhibited distinct sensor characteristics compared to both DC and FC.

4. Discussion

This study compared features derived from physiological signals recorded during different types of cravings in participants receiving BUP-XR for OUD. In this participant set, marked differences in accelerometer-based features were observed between drug cravings and food cravings. These features, which included time domain mean and skewness, spectral median power and area under the power density curve, and nonlinear sample entropy were higher during drug cravings. Another key finding was that mixed cravings which were simultaneously attributed by users to nicotine and food related items such as sugar, caffeine etc. differed significantly from both drug and food cravings. Higher values were observed during MC especially in heart rate and HRV features. Despite nicotine being an addictive agent, more features differed between mixed and drug cravings than between mixed and food cravings. This suggests that mixed cravings with both food and nicotine attributions by the participants, are not solely impacted by the presence of nicotine but also by the food, making them a unique type of craving. Interestingly, food and drug cravings had comparable heart rate and HRV features.

Machine learning models developed using PCA features performed well in distinguishing each type of craving relative to the alternative class comprising of the other two craving types. In the case of drug cravings, a maximum AUROC of 0.76 was achieved while mixed cravings were distinguished from the grouped food and drug cravings class with a maximum AUROC of 0.80. These findings suggest that physiological data carries differential information about craving types supporting our hypothesis. Following validation on larger data sets, physiological signal modeling can be leveraged to identify cravings by type, an application with practical interventional use in regulating OUD (and other SUDs).

Functional neuroimaging techniques, including functional magnetic resonance imaging (fMRI), positron emission tomography (PET), and single-photon emission computed tomography (SPECT), have consistently demonstrated that cues related to food, drugs, and alcohol activate key regions of the brain’s reward circuitry (Noori et al., 2016). While there is significant overlap in the neural regions engaged by these cues, distinct and discernible differences have been identified between areas activated during food cravings versus those activated during drug or alcohol cravings (Noori et al., 2016; Tomasi & Volkow, 2013). Consequently, individuals may understandably confuse food cravings with cravings for alcohol or other substances of misuse, as these sensations might subjectively feel similar (Czarnecki et al., 2023).

Given that there are neuroanatomical differences in the activation patterns elicited by food and drug cues, it logically follows that measurable physiological responses to these cravings may also differ. Supporting this hypothesis, studies have shown that individuals with SUD experience more intense neural activation in response to drug-specific cues compared to food-related cues (Berridge, 2024). If the physiological response magnitude correlates with neural activation intensity, then drug cravings should yield a more robust physiological signal than food cravings. This measurable differentiation provides a promising opportunity for objectively distinguishing drug cravings from food cravings.

Clarifying this distinction holds two major implications for our research. First, reliably differentiating food cravings from drug cravings would significantly enhance craving detection technologies’ accuracy and, therefore, clinical utility. Second, and critically, accurately identifying the type of craving experienced by individuals can guide tailored interventions. By providing patients with precise feedback regarding their cravings, interventions such as the RAE application (Carreiro et al., 2021) can offer timely and context-appropriate strategies, thereby potentially reducing the risk of return to drug use and supporting sustained recovery. This could also extend the use cases of such tools to other clinical states such as obesity and binge eating disorder.

This work has several important limitations. The analysis was carried out on a limited number of cravings and from a small group of participants and hence must be validated on a larger cohort. The cohort was also racially and ethnically homogenous, and all were receiving treatment with BUP-XR, which limits generalizability. Craving responses from individuals may vary based on their personalized baselines, type and number of substance(s) used, and this may be reflected in recorded sensor signals. This variability was not accounted for in this preliminary analysis which followed a population-based approach and could be overcome through the use of personalized models once more data are available. With the limited observation set, we resorted to dimensionality reduction prior to modeling, due to which feature importance of the ML models could not be examined. Finally self-reported cravings also carry some inherent bias, and cravings may be under-reported (which are precisely the reasons we need this tool).

Extended-release buprenorphine itself has an impact on sympathetic nervous system (SNS) activity and on cravings, so it is important to interpret the data in this context. Treatment with BUP-XR has been shown to decrease craving scores (Nunes et al., 2024), likely due to mu opioid receptor occupation. Alleviation of opioid withdrawal symptoms and anxiety may also decrease SNS activity (which is typically elevated in these states). Both of these factors may impact both the frequency of cravings and the detection capability using the RAE system and should be considered a limitation of this small pilot study.

Future research should focus on addressing some of our limitations with larger datasets, understanding other potential key craving subtypes, developing personalized models, and integrating these algorithms into intervention tools in clinically useful ways.

5. Conclusions

We present data to advance the current status of digital craving detection through the modeling of craving subtypes, which is a prime target for future research. Once refined, these models can be used to improve digital health treatments for SUDs (including OUD), food related health disorders, and general health and wellness applications.