Discriminating cocaine use from other sympathomimetics using wearable electrocardiographic (ECG) sensors

Abstract

Background:

Our group has established the feasibility of using on-body electrocardiographic (ECG) sensors to detect cocaine use in the human laboratory. The purpose of the current study was to test whether ECG sensors and features are capable of discriminating cocaine use from other non-cocaine sympathomimetics.

Methods:

Eleven subjects with cocaine use disorder wore the Zephyr BioHarness^™ 3 chest band under six experimental (drug and non-drug) conditions, including 1) laboratory, intravenous cocaine self-administration, 2) after a single oral dose of methylphenidate, 3) during aerobic exercise, 4) during tobacco use (N=7 who smoked tobacco), and 5) during routine activities of daily inpatient living (unit activity). Three ECG-derived feature sets served as primary outcome measures, including 1) the RR interval (i.e., heart rate), 2) a group of ECG interval proxies (i.e., PR, QS, QT and QTc intervals), and 3) the full ECG waveform. Discriminatory power between cocaine and non-cocaine conditions for each of the three outcomes measures was expressed as the area under the receiver operating characteristics (AUROC) curve.

Results:

All three outcomes successfully discriminated cocaine use from unit activity, exercise, tobacco, and methylphenidate conditions with a mean AUROC values ranging from 0.66 to 0.99 and with least squares means values all statistically different/higher than 0.5 among all subjects [F(3, 99) = 3.38, p =0.02] and among those with tobacco use [F(4, 84) = 5.39, p = 0.0007].

Conclusions:

These preliminary results support discriminatory power of wearable ECG sensors for detecting cocaine use.

1. Introduction

The standard of care for monitoring cocaine use poses significant limitations. Currently, approaches often rely on collection of urine samples several times weekly and subjective self-report susceptible to recall biases. In addition to methods used to adulterate results of urine toxicology testing, this technique exhibits poor time resolution/granularity such that the window for detection of cocaine metabolites is approximately 72 hours. A positive urine sample can reflect multiple use scenarios (i.e., a single use or multiples instances of use). Collecting urine samples several times weekly poses logistical challenges. Considering the collection of biological specimens typically necessitates patients to travel to clinician’s offices, it is typically not convenient, efficient, nor economical for monitoring progress. As such, poor time resolution or logistic barriers decrease monitoring compliance and hinder clinicians from conducting sufficient “in real time” drug use functional analyses and interventions. To overcome physical and temporal limitations of these traditional methods, there has been increased interest in optimizing tools available for cocaine-use detection (Barroso, Gallardo et al. 2009, Preston, Vahabzadeh et al. 2009, Preston and Epstein 2011, Epstein, Tyburski et al. 2014). For example, there has been work on increasing the number of biological samples (i.e., oral fluid or hair) available for detection of cocaine use (Barroso, Gallardo et al. 2009). While these approaches offer larger windows for detecting cocaine use, they remain limited by poor time granularity and vulnerability to adulteration (Wille, Baumgartner et al. 2014). Thus, instead of focusing on biological specimen collection, recent research has started to investigate mobile and wireless technology healthcare (mHealth) (Carreiro, Chai et al. 2017). MHealth may transcend logistic limitations by allowing remote assessments of patients (Carreiro, Chai et al. 2017). MHealth may also provide time resolution/granularity conducive to the delivery of preventive/therapeutic interventions at times when most needed (Carreiro, Chai et al. 2017), while lowering costs and boosting efficiency (Metcalf, Milliard et al. 2016).

MHealth offers multiple options including wearable devices or biosensors, handheld electronic diaries or smartphones, social media, telemedicine, and other tools. For example, ecological momentary assessments (EMAs), involving questions regarding their emotional states, geographic locations, current activities and other topics that subjects can answer while in possession of a handheld electronic diary, have shed light on situations preceding, triggering, and surrounding cocaine use. Such studies have revealed positive correlations between stress, craving, and cocaine use (Preston, Vahabzadeh et al. 2009, Preston and Epstein 2011), as well as physical locations associated with triggering emotional states (Epstein, Tyburski et al. 2014). However, similar issues of self-report remain, as subjects may not report every episode of cocaine use. When investigators use urine tests to assess for non-reported instances of use, limited temporal resolution of urine testing does not facilitate their precise identification. Thus, uncertainty remains as to whether non-reported episodes may be similar to or different from reported ones (Epstein, Willner-Reid et al. 2009).

Another area of MHealth pertains to wearable devices or biosensors, which can capture multiple physiological measurements such as electrodermal activity (EDA), electrocardiogram changes, skin temperature, and movement, among others (Sharma, Saeed et al. 2017). Importantly, several measurements are modified by use of cocaine and other drugs. For instance, one group demonstrated increases in EDA after parenteral administration of opioids or cocaine as analyzed from data collected via a wrist-worn device (Carreiro, Smelson et al. 2015). Another group tested changes in EDA assessed using ankle bracelets among veterans with post-traumatic stress disorder (PTSD) and substance use disorders. This study paired events of “arousal” or sympathetic nervous system activity based on EDA changes with phone-delivered therapeutic messages/interventions (Fletcher, Tam et al. 2011). As these “in-the-moment” wireless sensors have not received similar attention as other more common tools such as social media or web-based interventions (Mohr, Burns et al. 2013), basic questions, such as their sensitivity and specificity, remain unaddressed. Accordingly, the purpose of this work was to test remote wireless sensor networks (RWSNs) in distinguishing cocaine use from other conditions, based on electrocardiographic (ECG) changes. As a powerful cardiovascular stimulant, we thought that cocaine use might be best detected by remote ECG sensors, as this technology can closely monitor changes in cardiovascular activity in and outside of the laboratory.

To the best of our knowledge, few prior publications have addressed this topic. However, our group has been exploring the use of wearable ECG sensors for phenotypic characterization of cocaine-induced physiological and subjective states for over a decade (Natarajan 2013, Natarajan 2014, Natarajan, Angarita et al. 2016, Gullapalli and 2019). Our initial work included an inpatient human (n=6) laboratory experiment of intravenous cocaine-self-administration (Matuskey, Pittman et al. 2012) to test whether cocaine use could be detected using wearable ECG sensors (Natarajan 2013). Once participants were habituated to day-to-day activities of an inpatient research unit, they underwent a safety eligibility session (e.g., administration by the research nurse of 3 fixed escalating cocaine boluses of 8 mg/70 kg, 16 mg/70 kg, and 32 mg/70 kg) and self-administration (i.e., randomization to three different one-hour “binge” sessions in which they could self-administer each one of these doses for as many as 12 boluses/hour). Results demonstrated sensitivity of the RWSN device to detect cocaine use (from inpatient unit activity) with high Areas Under the Receiver Operator Curve (AUROCs) (Natarajan 2013). Further, results helped to establish the sensitivity/validity (e.g., dose-response relationship) of RWSN ECG features (i.e., a pattern showing higher sensitivity in response to higher doses) (Natarajan 2013). We then tested a different approach to extracting and labeling the morphological features of the ECG in order to address former limitations such as the presence of noise. These analyses showed that the new data processing pipeline could overcome initial methodological limitations (Natarajan 2014).

Next, we brought testing of RWSN devices from the human laboratory to the field (Natarajan, Angarita et al. 2016). We strived to overcome ecological constraints of testing wearable devices in the human laboratory setting as well as hindrances surrounding the identification of “ground true labels” in the field. Under laboratory conditions, the context surrounding cocaine use (e.g., precise timing of use and/or behavioral contexts) is clear and controlled throughout monitoring. However, in real-world settings, remote monitors may capture data influenced by several other factors in addition to cocaine use. Relying only on patients’ self-report for the context surrounding their cocaine use is not always a valid approach as simultaneous use of other substances (e.g., alcohol) and/or physical activity may remain unreported due to recall bias. Thus, we presented methods to address factors that may potentially influence lab-to-field generalizability of cocaine detection. In order to improve the validity of data surrounding when and in what context cocaine use may take place during monitoring, we explored the performance of frequent urine toxicology testing as well as the implementation of EMAs with smartphones (Natarajan, Angarita et al. 2016).

Here, we aimed to improve upon on our former efforts at detecting cocaine use with RWSN devices. The purpose of the current work was to explore whether the discriminatory power of remote wearable ECG sensors (i.e., a RWSN chest band) could be expanded from distinguishing cocaine use from unit activity (Natarajan 2013) to the discrimination of cocaine use from several “false positive” conditions: exercise, methylphenidate (MPH) administration, and tobacco use. To the best of our knowledge, there has been only one prior publication by another group testing the discriminatory power of wearable ECG sensors in the field for detection of cocaine use (Hossain 2014). This group tried to discriminate between cocaine use and physical activity using changes in the autonomic nervous system. Here, we present the novel inclusion of additional sympathomimetic potential false positive conditions in addition to incorporating sensors which measure cardiovascular changes (Hossain 2014).

2. Methods

2.1. Participants

The participants were 11 individuals with prior extensive cocaine use, in otherwise good medical, neurological and psychiatric health as confirmed by structured clinical interview for DSM-IV (SCID) (First 1995). All participants reported use of intravenous and/or smoked forms of cocaine (Table 1), and none were seeking treatment. All inclusion and exclusion criteria were identical to those of Matuskey et al. (2012).

Table 1.

Demographics

Demographic		Value / Mean (s.d.)
Sex (Male:Female)		9:2
Age		44 (6)
Years of Education		13 (2)
Race (#African-American: #Caucasian: # Native-American/African-American)		8:2:1
Cocaine Use
	Lifetime use (Years)	18 (8)
	Use per month (Days)	18 (9)
	Amount per use ($)	114 (82)
Cigarette Use
	Yes:No	9:2
	Cigarettes per day	8 (3)
Alcohol Use
	Days per month	4 (4)
	#Drinks per use	3 (2)

2.2. Study Procedures

2.2.1. Screening & Eligibility

The participants were screened for eligibility as outpatients, initially by telephone, and then, as appropriate, in person on the Clinical Neuroscience Research Unit (CNRU) of the Connecticut Mental Health Center (CMHC). In-person screening included semi-structured interviews, physical examination and laboratory work as described previously (Matuskey, Pittman et al. 2012).

Once participants finished screening and eligibility was confirmed, they were admitted to the CNRU, an elective 12-bed, locked inpatient psychiatric clinical research unit where visitors are restricted and access to drugs is prevented. Participants resided on the unit for up to 12 days and left the unit only to participate in experimental procedures, or for up-to-three-times-per day off-unit breaks (15–20 minutes each) under continuous staff supervision. The study was registered in the U.S. National Library of Medicine clinical trial database: NCT2018263. All study-related procedures were approved by the Yale University Human Investigation Committee (HIC) and performed in accordance with the Declaration of Helsinki. Participants provided voluntary written informed consent.

2.2.2. Device

As described previously (Natarajan 2013), raw ECG data were sampled using a Bioharness 3 chest band (Zephyr^™ Technology Corporation, Annapolis, MD; https://www.zephyranywhere.com) which utilizes RWSN technology. We focused exclusively on ECG data as collected by the Zephyr Bioharness chest band and communicated in real time to a Samsung Nexus smartphone via wireless/Bluetooth and subsequently via hardwire connection, to a secure, network-based server. From here on out in this work, we will refer to this device as the RSWN chest band.

2.2.3. Experimental conditions

Participants were monitored under several (up to six total) a priori experimental conditions designed to test whether the device could use data from ECG feature sets (see section 2.2.4) to discriminate cocaine use from other cardiovascular stimulants, including drug (e.g., pharmacologically related and unrelated stimulants) and non-drug (e.g., exercise or daily behaviors) activities. Specifically, subjects wore the RWSN chest band under conditions of 1) laboratory cocaine administration, 2) oral MPH administration, 3) aerobic exercise, 4) routine inpatient unit activity, and 5) tobacco use (in participants who smoked tobacco).

2.2.3.1. Cocaine administration:

All participants engaged in a well-validated single-day, 6-hour cocaine administration paradigm comprised of 3 consecutive phases: a baseline phase (i.e., during which no cocaine was administered), three nurse-initiated fixed-dose cocaine sessions, and three “binge” cocaine self-administration sessions (Natarajan 2013).

Fixed-dose phase:

This phase included 3 fixed-dose, nurse-initated cocaine sessions, each lasting 20 minutes. At the start of each session, participants received a single-bolus intravenous (IV) injection of cocaine using a fixed-order, ascending dose regimen of 8, 16, and 32 mg per 70kg respectively with a 100kg cap per adjusted dose. Physiological measurements, such as vital signs and cardiovascular assessments (i.e., ECGs), occurred at 5-minute intervals throughout each session. Accordingly, this phase was used to assess safety and eligibility for the “binge” self-administration paradigm which followed.

Self-administration “binge” phase:

The self-administration phase included 3 sessions where participants could self-regulate the amount of cocaine they received. Each session lasted 1 hour and utilized one dosage level (i.e., 8mg, 16mg, or 32mg), with the order of the dosage level randomized (i.e., participants could control how many times they self-administered the randomized dose within each session) (Matuskey, Pittman et al. 2012). Monitoring with the RWSN chest band occured throughout the course of this “binge” phase (i.e., during the three 1-hour sessions), allowing the device to be trained and tested with an ample amount of ECG data. Please see supplementary material 2 for more information on the number of data cases extracted for analysis.

2.2.3.2. MPH administration:

All participants (N = 11) underwent a 2.5 hour-long MPH administration session. The session consisted of one 45mg oral dose of MPH followed 1 hour later by 90 minutes of monitoring with the RWSN chest band. This timeframe falls in line with data on the pharmacokinetics of MPH which indicate peak plasma concentrations 2-hours post-administration (Candido, Menezes de Padua et al. 2021). Other physiological data (e.g., vital signs) were monitored clinically using procedures identical to those for the fixed-dose cocaine phase, except that subjects did not have intravenous catheters.

2.2.3.3. Exercise session:

All participants engaged in a session of aerobic exercise, including playing ping-pong [N=2] or riding a stationary bicycle [N=8]. The exercise session was 15 – 20 minutes in duration, and participants were monitored with the RWSN chest band throughout the entirety of the session.

2.2.3.4. Tobacco use:

Seven participants smoked cigarettes regularly, enabling the collection of ECG data during tobacco use. These subjects underwent monitoring with the RWSN chest band for a period of 5–7 minutes while actively smoking cigarettes during one daily break outside of the unit.

2.2.3.5. Unit activity:

Participants (N = 11) were monitored during activity on the inpatient unit. Specifically, they wore the RWSN chest band for approximately 30 minutes of unstructured elective time (i.e., participants were aware they were being monitored without explicit instructions regarding what to do). Participants did not have access to the previously mentioned aerobic activities, but could rest, read, play pool, watch television, use unit computers, or listen to/play music.

2.2.4. Data processing, feature extraction and primary outcomes

Raw ECG data were sampled at 250Hz using the BioHarness^™ 3 RWSN chest band. Three distinct ECG-derived feature sets served as primary outcome measures in assessing discriminatory power: 1) the RR interval measuring heart rate, 2) ECG interval proxies estimating PR, QRS, QT and QTc intervals, and 3) full ECG waveforms involving complete RR data traces. For more specific information regarding data processing and feature extraction refer to Supplementary Material 2. Data processing encompassed techniques developed in our preceding work (Natarajan 2013, Natarajan 2014).

2.2.5. Classification

Discrimination of cocaine use from non-cocaine experimental conditions occurred using a two-stage machine-learning algorithmic approach outlined in Supplementary Material 2. We defined discriminatory power according to AUROC values that reflect both sensitivity and specificity, with values ranging from 0.5 (reflecting no diagnostic capability or equally capable of producing a true positive as well as a false positive result) to 1 (reflecting a perfect test) (Bewick, Cheek et al. 2004). AUROC was a particularly suitable metric for this experimental design in relation to the imbalanced number of data cases available for analysis between conditions. In other words, the experimental conditions varied in length in order to emulate real-world settings, leading the amount of data cases extracted to vary between conditions as well. The AUROC metric takes into account class imbalances (i.e., the difference in data sample sizes between conditions), making it an ideal measurement of discriminatory power in this study (Bewick, Cheek et al. 2004).

2.2.6. Statistical analyses

Using AUROC values as dependent variables, we compared discriminatory powers of primary outcome measures (RR vs. ECG interval proxies vs. waveforms) for distinguishing cocaine use from other experimental conditions (discrimination tests) for each of the “within-subject” classifier trainings in stepwise statistical tests (i.e., among all participants first, followed by participants who smoked). AUROC levels were evaluated using a linear mixed model that included ECG-derived feature sets (RR, ECG interval proxies, and waveforms) and conditions (exercise, MPH administration, unit activity, and tobacco use [smoking group only]) as within-subjects factors and random subject effects.

To analyze discriminatory power based on ECG interval proxies, we used a composite ECG-derived feature set which combined data cases extracted from all ECG interval proxies instead of calculating discriminatory power based on individual ECG interval proxy sets. For example, discriminatory power based on PR interval proxy data alone was not analyzed. Rather, analysis occurred using a data set encompassing PR, QRS, QT, and QTc interval proxy data together. To our knowledge, literature on the impact of cocaine use on individual ECG intervals (e.g., PR, QRS, QT or QTc intervals) is still sparse and occasionally contradictory (Magnano, Talathoti et al. 2006) (Hollander, Lozano et al. 1994) (Daniel, Pirwitz et al. 1995). Thus, we suspected that cocaine’s manipulation of particular ECG intervals alone may not provide sufficient data to discriminate between conditions. Grouping ECG interval proxy data within a composite ECG feature set for analysis provided a significant signal for discriminating cocaine use from other conditions.

The correlation between repeated measures on an individual was modeled using random subject effects and structured variance-covariance matrices. The best-fitting model was assessed using information criteria. All multi-way interactions were modeled. Least square means and associated 95% confidence intervals were estimated from the model. All analyses were conducted using SAS, version 9.4 software (Cary, NC).

3. Results

Participant demographics are presented (Table 1). As classifiers were trained to analyze complexes with three peaks (P, R, and T) and two troughs (Q and S), one subject was excluded because of ECG complexes with six deflections (e.g., P, Q, R, S, R’, and T), thereby precluding comparisons of all outcome measures (including ECG interval proxies) in all 11 subjects. In two subjects, documentation of smoking onset/offset was unclear and occurred in the context of other potentially confounding activities (e.g., walking outside to/from a smoke break). As a result, data from these two participants were not included in tobacco discrimination analyses.

3.1. Areas Under the Receiver Operator Curve (AUROC):

Figure 1 presents classifier testing using the within-subject (“personalized”) training approach. Specifically, mean (± standard error [SE]) AUROC values are depicted for each of the three outcome measures (RR [red], ECG interval proxies [orange], and waveforms [yellow]) across three discrimination tests (i.e., cocaine vs. exercise, MPH, and unit activity) for all participants (N=10). We found high levels of discriminatory power for all three outcome measures across all three discrimination tests, with all AUROC curve values statistically different from 0.5 [F(3, 99) = 3.38, p = 0.02] and ranging from 0.71 (using RR to discriminate between cocaine and MPH) to 0.99 (using waveforms to discriminate between cocaine and unit activity) (Figure 1).

Figure 1. Discriminatory Power of ECG Feature Sets Among All Conditions in All Participants.

Least Square (LS) means of AUROC (+/− Standard Error [SE]) for RR, ECG interval proxies, and full waveforms among all participants (N = 10) for the following discriminations: cocaine vs. exercise, cocaine vs. methylphenidate, and cocaine vs. unit activity.

Figure 2 presents classifier testing using the within-subject (“personalized”) training approach for participants who smoked tobacco (N=7). Mean (±SE) AUROC curve values are depicted for RR [red], ECG interval proxies [orange], and waveforms [yellow] across four discrimination tests (cocaine vs. exercise, MPH, unit activity, and tobacco). These results also revealed high discriminatory power with all AUROC values statistically different from 0.5 [F(4, 84) = 5.39, p = 0.0007] (Figure 2).

Figure 2. Discriminatory Power of ECG Feature Sets Among All Conditions in People who Smoke Tobacco.

Least Square (LS) means of AUROC (+/− Standard Error [SE]) for RR, ECG interval proxies, and full waveforms among people who smoke tobacco (N = 7) for the following discriminations: cocaine vs. exercise, cocaine vs. methylphenidate, cocaine vs. unit activity, and cocaine vs. tobacco.

There were no interaction effects between outcome measures and discrimination testing among all subjects; thus, post-hoc comparisons were not conducted. However, there was an interaction effect among participants who smoked tobacco [F(8, 84) = 2.53, p = 0.02]. Post-hoc explorations revealed that waveforms outperformed ECG interval proxies and RR on discriminations between cocaine and MPH, and cocaine and tobacco. Furthermore, ECG interval proxies outperformed RR on discriminations between cocaine and MPH, unit activity, and tobacco. See Supplementary Material 1 for specific values relevant to these post-hoc tests.

3.2. Device comfort & use feasibility:

Throughout monitoring with the RWSN chest band (i.e., the Zephyr^™ BioHarness 3), research staff checked in with participants regarding the comfort of the device. No complaints were noted, likely due to the adjustable nature of the chest band, which can be tightened or loosened according to the patient’s preference. Participants often expressed curiositiy about how the device worked, checking in with staff to be sure the device wasn’t too loose or being worn incorrectly. It is important to note that a balance needed to be struck between the level of comfort of the device and the ability to most accurately record data. If the band was too loose or became loosened over the monitoring session, it may not have captured all data. Thus, it was necessary to explain to the patient how the chest band should be properly worn and to check that it was fit sufficiently close to the patient’s chest for accurate data collection.

4. Discussion

4.1. Summary

To our knowledge, this is the first study to explore whether ECG signals from wearable remote wireless sensors can discriminate between cardiovascular effects of cocaine and those produced by other cardiovascular stimulants (including MPH, aerobic exercise, and tobacco). Our results show excellent levels of discriminatory power for within subjects’ analyses, both among all participants and those who smoked tobacco. This level of sensitivity/specificity is evidenced by high AUROC values for all three ECG feature sets and among all discriminations. Our results also reveal that ECG waveforms exhibit higher discriminatory power in comparison to ECG interval proxies, and that ECG interval proxies demonstrate higher discriminatory power than heart rate among individuals who smoked tobacco.

4.2. Interpretations

One initial hypothesis for the high specificity of all three ECG feature sets within subjects could be that their discriminatory power reflects differences in heart rates across experimental conditions. However, considering ECG interval proxies and waveforms generally exhibited higher discriminatory power than RR, it seems likely that discriminatory power is influenced by nuances beyond differences in heart rate. The physiology of cocaine’s cardiovascular effects, in comparison to those of MPH, tobacco, and exercise, may explain both the overall specificity of the RWSN and differences between the three ECG feature sets’ discriminatory performances.

Aside from blocking dopamine transporters, cocaine can inhibit voltage-gated sodium channels (I_Na) in a manner similar to quinidine or other class I antiarrhythmics, an effect that has been associated with QRS prolongation (Ramirez, Femenia et al. 2012). Similarly, cocaine can inhibit rapid delayed rectifier potassium currents (I_Kr), which may result in QT prolongation and can both inhibit and stimulate L-type calcium currents (Ramirez, Femenia et al. 2012). ST elevation has also been reported following cocaine administration, even in the absence of symptoms (Ramirez, Femenia et al. 2012). Reports of cocaine’s effects on the PR interval have been more heterogeneous, perhaps due to differences in methodologies including circumstances surrounding ECG measurement and timeframes between exposure and measurements. While several case reports indicate cocaine-induced PR prolongation (Ramirez, Femenia et al. 2012), some results were in the contexts of other factors such as acidosis (Kalimullah and Bryant 2008) or alcohol use (Weiner, Weiner et al. 2008). Other clinical studies have produced conflicting reports, such as shortening of the PR interval after inhalation of cocaine (Magnano, Talathoti et al. 2006), no differences in the PR interval between people who used cocaine and those who did not (Hollander, Lozano et al. 1994), and no PR changes after intranasal administration of cocaine (Daniel, Pirwitz et al. 1995). While more work should be done to further investigate specific effects of cocaine on each ECG interval, cocaine administration may lead to a specific ECG signature. Given cocaine’s effect not only on the duration but height of different intervals, it follows that a more comprehensive approach, such as the full ECG waveform, may display higher discriminatory power than less inclusive measures.

Like cocaine, other stimulant medications block the reuptake of dopamine and norepinephrine into the presynaptic neuron, however, without the same effects on Na/K channels. Thus, we may infer that these substances may have ECG signatures different from cocaine’s. Indeed, no strong data suggest that MPH prolongs QTc intervals (Stiefel and Besag 2010). In fact, some research has suggested that cardiovascular effects including QRS and QT intervals post-MPH-overdose do not differ significantly from other non-cardiotoxic medications, like acetaminophen (Hill, El-Khayat et al. 2010). This suggests that cocaine-induced ECG changes may be detectable and distinguishable from those of other sympathomimetics, like MPH.

Similarly, tobacco may exhibit cardiovascular effects distinct from cocaine’s. While tobacco’s cardiovascular effects are typically ascribed to nicotine enhancing release of catecholamines following binding to nicotinic cholinergic receptors (nAChRs) in different structures like the adrenal medulla or autonomic ganglia (Benowitz 1988), nicotine may also modulate other cardiac channels (Wang, Yang et al. 2000). For instance, nicotine has demonstrated inhibition of cardiac A type potassium channels like the transient outward K (+) current (I:(to)) and Inward Rectifier Potassium Channels (K_ir )(Wang, Yang et al. 2000). Although some effects have been documented by studying canine myocites, limiting applicability of findings to humans, it is interesting that these effects remain after removing potential influences of catecholamines (Wang, Yang et al. 2000). Accordingly, ECG results from a group of people who smoked tobacco (N = 44) 2 hours post-cigarette compared to those from people who did not smoke tobacco revealed non-statistically significant QTc shortening and statistically significant shortening of RR intervals, ST segments, and QT intervals (Devi, Arvind et al. 2013). These results support the hypothesis that nicotine could produce significant ECG changes (Wang, Shi et al. 2000) aside from tachycardia (Wang, Yang et al. 2000), facilitating its distinction from cocaine. However, as tobacco has many active chemical moieties, further research is needed into the precise mechanisms underlying the current findings.

In terms of physical activity, aside from increased heart rate, exercise can lead to other ECG changes (Simoons and Hugenholtz 1975) like depression of QRS-ST junctions (Sjostrand 1950), increased P-wave amplitudes (Irisawa and Seyama 1966, Simoons and Hugenholtz 1975), shortened intervals between the onset of QRS and maximum spatial amplitudes of T waves (Simoons and Hugenholtz 1975), depression of ST segments, and decreased T wave amplitudes (Deckers, Vinke et al. 1990). Akin to MPH or nicotine, the above literature supports the capacity of exercise to produce ECG changes distinct from cocaine-induced ones.

Potentially, conditions such as MPH, tobacco, or exercise may lead to false positives. However, examining the entire ECG trace involving full waveforms without a priori assumptions may exhibit higher discriminatory power than approaches using a priori assumptions about relevant ECG features, such as ECG interval proxies. Such data also suggest why both approaches exhibit higher discriminatory power than heart rate, as this approach is based on only one feature and does not capture additional complex and more subtle changes related to other conditions.

4.3. Limitations

The external validity of this work is limited by several factors. First, there was a relatively small sample size and unbalanced sex ratio among participants. The study sample included only two females, both of whom smoked, as opposed to the 7 males included who varied in smoking status (note: the two non-smoking individuals were male). Considering that some research has suggested that the cardiovascular responses of self-administered cocaine may vary between sexes, a very necessary modification to the study design moving forward would be to include more females (some of whom smoke, and some of whom do not) (Lynch, Kalayasiri et al. 2008).

The work was also constrained during the process of data collection as a result of obstacles with algorithm training/data processing. Some data could not be analyzed due to an extra deflection found for ECG complexes. This phenomenon rendered the logistic classifier inapplicable, even though it can be normal to have an extra R (known as R’), or other deflection modifications, depending on the individual and/or lead placement.

Our sole use of intravenous cocaine administration likely limited the external validity of the work as well. Outside of the laboratory, smoked and intranasal cocaine use are more common routes of administration than using cocaine intravenously (Paczynski 2011). Simultaneously, because only an IV cocaine self-administration paradigm was approved for this work, we did not include participants whose primary cocaine route of administration was intranasal. This criterion was included to avoid exposing individuals to a route of administration which may be more addictive than their current route (i.e., smoked and IV cocaine may be more addictive than intranasal cocaine use) (Roque Bravo, Faria et al. 2022).

A final limitation is inherent in the uneven heart rates generated by the different conditions, as it is possible that discriminatory power was being driven by these differences as opposed to other more subtle ECG changes. This limitation can be partially addressed by an analysis which matches for heart rates across conditions and within subjects. However, this may significantly reduce the amounts of ECG data cases available for analysis, and could potentially select outlier ECG templates from each condition (e.g., templates on the lower end of the spectrum of heart rates generated by the cocaine session, or templates on the upper end generated by MPH).

4.4. Future Studies

One significant way this work may be extended will include training algorithms to define personally individualized baseline ECG status, which may have a variety of components different from 3 peaks (P, R, T) and 2 troughs (Q and S). Once algorithms are trained on personal normal ECGs, they can be tested to identify changes induced by cocaine or other conditions. Such studies could also implement new methodologies for peak extraction (Natarajan 2014), addressing limitations surrounding classifier training.

Future research should account for varying cocaine routes of administration (ROAs) as well. For instance, to our knowledge, there is currently no literature which directly compares differences between the specific cardiovascular impact of intravenous, intranasal, and smoked cocaine. However, there are known differences in pharmacokinetics between ROAs. The physiological effects of instranasal cocaine use have been documented to occur much later than those of smoked or IV cocaine use (i.e., in order to reach peak plasma concentration, intranasal cocaine may take up to 10 times longer than smoked or IV cocaine, whose peaks occur similarly at around 5 minutes post-administration) (Roque Bravo, Faria et al. 2022).

Different ROAs may also facilitate the intake of more or less cocaine over time. Using cocaine intravenously could allow a faster intake of high cocaine doses over a shorter period of time than what is possible with other ROAs (Wise and Kiyatkin 2011). Considering that our previous research suggested that the RWSN chest band’s sensitivity to ECG changes varies in relation to cocaine dose, it may be beneficial for future studies to assess the device’s “tipping point” for detection of cocaine use. In other words, research across several cocaine ROAs should investigate the temporal point at which the discriminatory power of the device becomes significant following cocaine administration. We suspect that with intranasal cocaine use, an ideal monitoring period may last longer or begin later than with other routes. This is because, as noted, unlike with IV or smoked cocaine whose peak plasma concentrations occur very quickly post-administration (i.e., after 5 minutes), intranasal cocaine use does not generate peaks in plasma concentrations until around 50 minutes post-administration (Roque Bravo, Faria et al. 2022).

Similarly, in order to eventually conduct outpatient studies with remote cocaine use monitoring in natural environments, it will be important to first test this device with a human laboratory paradigm of smoked cocaine. Despite their similiarities in addictiveness, cocaine use via IV administration (i.e., the ROA utilitized in this work) and smoked cocaine use (i.e., a more common cocaine ROA) still differ in terms of metabolization and thus, acute cardiovascular impacts (Roque Bravo, Faria et al. 2022). Significant experience with smoked cocaine use exists in several laboratories (Haney, Hart et al. 2005, Hart, Haney et al. 2008, Dakwar, Hart et al. 2017) so it may be feasible to conduct research to test the RWSN chest band’s discriminatory power within this context.

In order to address the study’s potential limitation involving the influence of heart rate differences across conditions, future work can train and test classifiers in ECG segments/traces with heart rates matched within-subjects across conditions. Based on our current results, we predict that these analyses will exhibit AUROC values close to 0.5 for RR, as the device would no longer be able to use the RR feature set to discriminate conditions for which heart rate was the same. This concern may also be addressed by including an MPH session in the study with administration of a 0.3 mg/kg intravenous dose, in order to attain experimental conditions ending with heart rates matching those generated by cocaine use. Our group has used this protocol in the past without compromising subject safety (Li, Morgan et al. 2010). Accordingly, sympathomimetic pharmacological agents apart from MPH may be included for this purpose. Stimulants like amphetamine and methamphetamines, methcathinone (a bath salt drug), and N-methyl-1-(3,4-methylenedioxyphenyl)propan-2-amine (i.e., MDMA or ecstasy) seem to produce cardiovascular responses similar to cocaine (Hossain 2014). Albeit outside of the initial scope of this work, administration of one or several of these agents could represent additional conditions which may be included in a follow-up study utilizing a heart rate-matched analysis.

Eventually, we hope to bring this work beyond the human laboratory, studying patients in the community and natural environments. For instance, we could algorithmically train devices on data gathered during inpatient laboratory procedures and then test them with data collected from study participants outside of the lab. This design would be an appropriate follow-up for our current findings which demonstrate high AUROC values within subjects, but not between subjects, and might further move this field in the direction of individualized medicine.

4.5. Importance of work

The most promising ultimate direction of this work is to to arrive at a point where RWSN devices can become not only tools for real-time, in-the-field detection of cocaine use, but also an avenue for preventive and therapeutic interventions. For example, some researchers have attempted to pair individualized information about estimated blood alcohol concentration with a therapeutic intervention which warns individuals about further alcohol consumption (Gajecki, Berman et al. 2014). Future studies could conduct similar work by first determining whether there are identifiable ECG changes preceding cocaine use via comparisons between unit activity and pre-cocaine baselines, or the period of time immediately preceding cocaine self-administration which indicates when subjects are anticipating cocaine use. If such signatures are detectable, this period may be targetted for preventive interventions with high clinical value. Equally salient is the possibility of offering therapeutic interventions in real-time. Replication of findings in outpatient or naturalistic studies could introduce future clinical studies testing a combination of cocaine use anticipation detection and detection of cocaine use, with preventive or therapeutic interventions, respectively.

Further, because these wearable devices may communicate with smartphones or other mHealth-related devices, this work may be expanded to integrate instances of drug use with relevant behaviors, affects, and contexts (Boyer, Smelson et al. 2010, Boyer, Fletcher et al. 2012). Being able to integrate unique aspects of each person’s drug use patterns (ECG changes, triggering locations or stimuli, etc.) could theoretically enhance sensitivity both in detecting cocaine use and cocaine use anticipation. All tools combined could create a fuller assessment, leading to personalized treatment (Carreiro, Chai et al. 2017). Ideally, the use of RWSN devices will be combined with other technological advances in the area of mHealth, offering patients the option to carry with them smartphones, heart-rate monitors, accelerometers, or other devices that could inform clinicians of individuals’ states in real time. Implementation of such approaches could foster connections between patients and providers such that medical care could be based both on personalized data and real-time input, reflecting an optimal understanding of the dynamic features contributing to cocaine-use-disorder symptomology and recovery.

Highlights.

• Electrocardiographic (ECG) changes occur with cocaine use.

• On-body ECG sensors can discriminate cocaine use from non-cocaine conditions.

• ECG-derived features exhibit varying degrees of specificity.

• Cocaine may exhibit ECG signatures unique from those of other sympathomimetics.

• Remote wireless sensors are promising mHealth tools for cocaine use detection.