Use of Vital Records to Improve Identification of Suicide as Manner of Death for Opioid-Related Fatalities

Abstract

Background:

Accurate classification of intentional death as suicide is essential to target prevention measures appropriately. Unfortunately, manner of death (MOD) for many opioid-related fatalities are unclassified in the United States, and in Maryland, as many as 82% of overdose deaths are classified as undetermined manner.

Aims:

For opioid-related fatalities in Maryland, leverage death certificate data to develop a model for identifying suicide as MOD among those classified as undetermined by the medical examiner.

Method:

Demographic and toxicology data were used to develop a classification model for opioid-related deaths where MOD was known, and then applied to a cohort where MOD was undetermined to estimate the likelihood that the intent was suicide.

Results:

Antidepressants, neuroleptics, oxycodone, benzodiazepines, and acetaminophen were more common in deaths classified as suicide while fentanyl, cocaine, and morphine were more common among accidental deaths. Our classification model correctly identified suicide cases 82% of the time (PPV = 0.82; AUC = 0.90) and expanded the number of suicide cases by 43% when applied to undetermined deaths.

Limitations:

The accuracy and completeness of death records.

Conclusions:

Data from standard autopsies can be used to detect additional suicide deaths with good statistical precision. Incorporating clinical information could enhance predictive accuracy and improve classification.

In the United States, suicide and opioid-related overdose deaths have been increasing for decades. In 2022, over 49,000 individuals died by suicide annually (14.2 deaths per 100,000; CDC, 2024a) and there were about 82,000 opioid-related deaths (32.6 per 100,000; CDC, 2024a, 2024b). Of course, these two groups are not mutually exclusive as many opioid-related deaths are also suicides. Several studies have shown that many deaths characterized as accidents or as deaths of undetermined intent are in fact likely suicides (Björkenstam et al., 2014; Kapusta et al., 2011; Liu et al., 2020; Marusic et al., 2003; Rockett et al., 2014; Stone et al., 2017; Varnik et al., 2010). Drug poisonings are particularly notorious for classification challenges (Breiding & Wiersema, 2006; Lachaud et al., 2018; Pamer et al., 2008; Rockett et al., 2010). It is estimated that 30–75% of overdose deaths classified as accidental or undetermined are in fact suicides (Donaldson et al., 2006; Liu et al., 2020; Pamer et al., 2008; Rockett et al., 2018).

While Maryland has the 5th lowest rate of suicide relative to other states at 9.5 deaths per 100,000 (CDC, 2024c), the statewide opioid-related fatality rate is higher than the national average at 36.1 deaths per 100,000 (CDC, 2024d). Although the rate of suicide in Maryland is relatively low as compared to other states, there is a likelihood that this rate may be underestimated due to manner of death (MOD) being classified as accidental or undetermined. The Office of the Chief Medical Examiner of Maryland (OCME) is conservative in categorizing intentionality among overdose deaths, leaving as many as 82% of overdose deaths classified as undetermined manner (Rockett et al., 2015a, 2015b), implying robust precision but poor sensitivity in classification of suicides and accidental deaths. Determining intentionality among drug overdoses is nuanced and involves a thorough death scene investigation, review of medical history, complete postmortem examination, including comprehensive toxicological testing, and combination of psychological and contextual factors, accounting for overlapping risk factors and limited information available at the scenes of death (Nestadt & Bohnert, 2020). Furthermore, there is great variation in the way manner (intentionality) of death is characterized, depending on examiner training and local protocols (Neeleman & Wessely, 1997). However, accurate classification of suicide is essential if we are to understand the differential risk of suicide across a population, to identify those most at risk, and allocate resources appropriately (Stone et al., 2017). Maryland has one of the country’s highest opioid overdose rates and maintains the largest proportion of overdoses classified as undetermined manner (CDC, 2024c, 2024d; Rockett et al., 2015a, 2015b, 2021), despite the death investigations being among the most comprehensive in the United States (Nestadt et al., 2017).

In this study, we investigate the magnitude and direction of possible MOD errors and identify factors which may indicate undercounted suicide using Maryland medical examiner’s data. Specific aims include (1) describe demographic and toxicology characteristics of Maryland opioid overdose-related fatalities, (2) identify toxicology and demographic features that can be used to retrospectively identify suicide as a manner of death, and (3) assess model performance in the context of unlabeled cases, and prospects of re-classifying undetermined deaths. We focus specifically on toxicology and demographic features since complete data are more likely for these characteristics within death certificates than other features such as psychological and contextual factors. Maryland is uniquely suited for this study, with one of the country’s most respected medical examiner systems that utilizes state-wide protocols and collects more complete decedent information than most other systems. Yet, Maryland also sustains the nation’s highest proportion of overdoses of undetermined intent (Rockett et al., 2015a, 2015b). The ability to better distinguish and understand these two interrelated crises is essential to targeting interventions for their prevention.

Methods

Data Source

Data were provided by OCME, a statewide medical examiner system which provides systematic, protocol-guided evaluations (inclusive of autopsy and toxicology) of all decedents who require investigation for the possibility of suicide, homicide, or accidental deaths. This includes all overdose deaths that occur in the State of Maryland. The OCME determines manner of death with a high degree of precision, leaving cases with any doubt to be characterized as undetermined when there is a lack of sufficient evidence to definitively classify death as natural, accidental, suicidal, or homicidal.

Study Design

A retrospective cohort study design leveraging demographic and toxicology data from medical examiner data was used to develop a classification model for suicide as manner of death. Study procedures included (1) application of descriptive statistics to characterize opioid-related overdose deaths stratified by the OCME assigned manner of death, (2) developing a classification model using a subcohort of deaths where MOD was classified as either accidental or suicide, and (3) applying the model to the subcohort whose manner of death was classified as undetermined to estimate the likelihood that the intent was suicide. This study was reviewed and granted waiver of consent to use records by the Maryland State Department of Health and the Johns Hopkins School of Medicine Institutional Review Board (Protocol #00228807).

Study Sample

The study sample included 10,218 records of individuals who died in Maryland from opioid-related overdose between January 1, 2016, and December 31, 2020. Overdose was determined as the cause of death by the medical examiner. We considered opioids a contributing factor if the decedent had a positive toxicology for any of the following opioids or their derivatives: fentanyl, heroin/morphine, or oxycodone. Decedents positive only for methadone or suboxone, opioid substances used to treat addiction, in the absence of other narcotics, were excluded. Deaths ruled naturally caused or homicides were excluded, leaving only cases where the manner of death was classified as either accidental, suicide, or undetermined. The time period was selected to ensure sufficient cases could be identified as having similar testing and environmental conditions, since the availability of drugs (such as fentanyl) has changed in the State of Maryland during this time.

Study Variables

Independent variables were sourced from OCME death certificate and toxicology report data files and included (1) demographic characteristics of age, sex, race/ethnicity, and county (used to construct National Center for Health Statistics (NCHS) urban/rural classification schemes); (2) toxicity data for opioids including fentanyl, heroin/morphine, and oxycodone and their associated derivatives of hydrocodone, hydromorphone, and oxymorphone; and (3) toxicity data for other substances known to be commonly causative or associated with overdose death in Maryland during this time period including cocaine, antidepressants, benzodiazepines, neuroleptics, ethanol, and acetaminophen.

For all toxicology data, variables were created as follows: First, binary class variables (0/1) were developed to reflect the presence of any substance within a particular drug type class. For example, Class: Antidepressant Positive indicated the presence of any antidepressant within the decedent’s toxicology report. Next, for specific substances within a drug class type where the amount present was quantified, two variables were created: (1) binary substance positive variables were developed to indicate the presence of specific substances within a drug type class (e.g., Substance: Bupropion Positive indicative of the presence of bupropion, a common antidepressant) and (2) binary substance high variables were developed to indicate high levels of specific substances within a drug type class (e.g., Substance: Bupropion High). To determine the threshold of high levels for a given substance, a sensitivity analysis was performed that leveraged decedent records where suicide and accidental manners of death were recorded as truth labels. For each substance, a series of classification metrics were calculated at different values for blood concentration, to delineate high versus low concentration. Establishing the appropriate cut-off was nuanced, but we sought to label cases where there were at least five high observations for blood concentration that yield a positive predictive value of 0.8 or greater to maintain very good precision in correctly identifying cases where the manner of death is suicide.

Among records in the study sample where manner of death was classified as either accidental or suicide, a binomial dependent variable (0 = accidental, 1 = suicide) was constructed to differentiate MOD for classification model development using regression procedures.

Statistical Methods

Descriptive Statistics of Sample Characteristics

Descriptive statistics, including frequencies of demographic and toxicology characteristics of the overall cohort as well as subgroups defined by MOD were calculated along with associated likelihood ratio χ² or Fisher exact test statistics. All analyses were conducted in SAS version 9.4.

Classification Model Development

A subcohort of decedent records where MOD was designated as either accidental or suicide by the Maryland OCME was used to develop the classification model to differentiate suicide from accidental deaths. A competing model configuration was used where class positive, substance positive, and substance high variables, along with the demographic variables of sex and age group (i.e., <25, 25–39, 40–64, 65+), where all evaluated for logistic regression model inclusion, and then stepwise backward elimination was applied to derive the final model. Two demographic variables, race/ethnicity and NCHS urban/rural designations, were ultimately not considered for model inclusion to avoid inadvertent bias in model predictions, but useful as descriptive characteristics of our cohort.

Model performance metrics were evaluated according to a moving decision threshold for response probability rather than percentile of risk, to illustrate differential effect on precision and recall. For example, a setting of > 0.5 reflects the probability of suicide being more likely than not, according to the model, while a setting of 0.8 would reflect a more stringent level of classification (i.e., greater precision, lower sensitivity). Results corresponding to multiple thresholds are provided but costs associated with misclassification are not equal across class of outcome. In short, to make use of the final model, we recognize a high degree of precision is needed for medical examiners, and therefore, higher thresholds for risk are preferred.

Model Application

Following model development, the classification model was applied to the subcohort where MOD was undetermined to yield probability estimates of each overdose death being due to suicide (vs. accident). Decedents in the undetermined cohort who had a probability estimate greater than 0.5 (i.e., p(x) ≥ 0.5) of having died by suicide (e.g., reclassified cases) were quantified and compared with decedents classified as suicide by OCME using likelihood ratio χ² or Fisher exact test statistics for individual attributes or substances.

Results

Demographic characteristics and toxicology results for the study cohort are presented in Table A1 in the Appendix at the end of the article. Of 10,218 unique decedent records in the study cohort, 134 (1.4%) were classified as suicide, 2,312 (22.5%) were classified as accidental, and 7,772 (76.1%) were classified as undetermined for MOD. Relative to accidental [a] and undetermined [u], suicide [s] cases were more likely to be female (55%_s vs. 27%_a/u; p < .0001), 65 years and older (26%_s vs. 4%_a/u; p < .0001), and White (87%_s vs. 61%_a/u; p < .0001). Accidental and suicide deaths were most commonly observed within large central and fringe metropolitan areas (77%_a and 85%_s, respectively) whereas undetermined deaths were most commonly observed in large fringe and medium metropolitan areas (75%_u; p_NCHS < .0001). Although some statistically significant differences were observed, the toxicology profile of accidental and undetermined deaths tended to be fairly similar to each other and distinctly different from suicide decedents for the following: Fentanyl, cocaine, and methadone were far more common in accidental (83%, 38%, 10%, respectively) and undetermined (89%, 39%, 8%) cohorts relative to suicide (21%, 9%, 3%), while prescription and over-the-counter drugs including oxycodone, antidepressants, neuroleptics, benzodiazepines, and acetaminophen were much more common among the suicide cohort (63%, 50%, 13%, 29%, 30%, respectively) compared to the accidental (15%, 21%, 6%, 16%, 7%) and undetermined (10%, 20%, 6%, 14%, 6%) cohorts. Testing positive for heroin/morphine was highest among undetermined MOD decedents (45%), followed by accidental (41%) and then suicide (37%) decedents. While statistically different, ethanol prevalence among the groups was similar, ranging from 35% among undetermined decedents to 31% among accidental decedents.

Performance metrics for the classification model suicidal intent among opioid overdose decedents are presented in Table 1. Notably elevated odds ratios for suicide compared to accidental MOD were observed for the variables oxycodone high (OR 16.8, CI 5.8–48.4) and acetaminophen high (OR 8.5, CI 1.8–40.2) and moderately elevated significant odds ratios were observed for hydrocodone positive (OR 3.4, CI 1.5–7.9), antidepressant positive (OR 3.3, CI 2.0–5.4), female sex (OR 2.1, CI 1.3–3.3), and age 65 and older (OR 2.0, CI 0.8–5.1). Although exceptionally elevated odds ratios for suicide versus accidental MOD were observed for high levels of trazodone (OR 147.5, CI 16.0–undefined) and bupropion (OR 22.0, CI 5.1–95.8), the wide confidence intervals are suggestive of estimate imprecision due to the low counts of these substances. Among odds ratios indicative of a greater likelihood of accidental MOD (i.e., OR < 1.0), the presence of fentanyl was the most pronounced with an odds ratio value of 0.1 (CI 0.1–0.2). The presence of the antidepressants citalopram (OR 0.4, CI 0.2–0.9) and trazodone (OR 0.2, CI 0.1–0.9) was also associated with greater likelihood of accidental MOD, as was ages 25–39 years (OR 0.2, CI 0.1–0.5) and 44 to 64 years (OR 0.3, CI 0.1–0.8) when compared to age less than 24 years.

Table 1.

Classification model to differentiate suicide from accidental intent among Maryland decedents^a who died from overdose involving opioids between 2016 and 2020

Variable	β	SE β	Wald χ2	p	OR	LCL	UCL
Constant	3.3524	0.8062	17.29	<.0001
Sex (female vs. male)	0.3606	0.1175	9.42	.002	2.1	1.3	3.3
agecat4 (25–39 vs. <24)	−1.0832	0.2269	22.78	<.0001	0.2	0.1	0.5
agecat4 (40–64 vs. <24)	−0.5946	0.1815	10.74	.001	0.3	0.1	0.8
agecat4 (65+ vs. <24)	1.1734	0.2492	22.18	<.0001	2.0	0.8	5.1
acetaminophen_high 1 versus 0	1.0727	0.3954	7.36	.007	8.5	1.8	40.3
benzoylecgonine_pos 1 versus 0	−0.4923	0.1929	6.51	.01	0.4	0.2	0.8
fentanyl_class_pos 1 versus 0	−1.0549	0.128	67.88	<.0001	0.1	0.1	0.2
hydrocodone_pos 1 versus 0	0.6128	0.215	8.12	.004	3.4	1.5	7.9
oxycodone_high 1 versus 0	1.4105	0.27	27.28	<.0001	16.8	5.8	48.4
ad2016_class_pos 1 versus 0	0.5971	0.1274	21.96	<.0001	3.3	2.0	5.4
citalopram_pos 1 versus 0	−0.4927	0.2219	4.93	.03	0.4	0.2	0.9
trazodone_pos 1 versus 0	−0.6968	0.32	4.74	.03	0.2	0.1	0.9
bupropion_high 1 versus 0	1.5461	0.3751	16.99	<.0001	22.0	5.1	95.8
trazodone_high 1 versus 0	2.4971	0.5669	19.40	<.0001	147.5	16.0	>999.9

Note.

^aBased on sample of 2,446 decedents among which MOD classified as accident for 2,312 decedents and suicide for 134 decedents.

LCL = lower confidence interval. OR = odds ratio. UCL = upper confidence interval. SE = standard error.

Table 2 reflects classification metrics for our model using various threshold settings. At a threshold probability of 0.5, the model captures 54 of 134 (40.3%) of suicide cases with 81.8% of predictions being correct (F1 32.3). A threshold of 0.7 results in 42 of 134 (31.3%) suicide cases being correctly detected, and 89.2% of all predictions being correct (F1 26.6). The area under the receiver operating characteristic (AUC) reached 0.9, indicating excellent discrimination ability of the classifier across various threshold values. The AUC depicted in Figure 1 shows different labeled probability points, demonstrating the classifier’s performance in terms of true positive rate and false positive rate at different thresholds. Notably, at 0.1, the sensitivity of the classifier reached 75%, indicating an ability to correctly identify a substantial portion of suicide instances, while maintaining a specificity level above 90%, highlighting the model’s capability to effectively avoid false positives.

Table 2.

Sensitivity settings for different decision thresholds of the classification model and subsequent performance metrics

	Correct		Incorrect		Percentages
Prob level	Event	Nonevent	Event	Nonevent	F1	Sensitivity	Specificity	PPV	NPV
0.3	70	2,258	54	64	35.7	52.2	97.7	56.5	97.2
0.5	54	2,300	12	80	32.3	40.3	99.5	81.8	96.6
0.6	44	2,302	10	90	27.3	32.8	99.6	81.5	96.2
0.7	42	2,307	5	92	26.6	31.3	99.8	89.4	96.2
0.8	28	2,307	5	106	18.6	20.9	99.8	84.8	95.6
0.9	20	2,308	4	114	13.7	14.9	99.8	83.3	95.3

Note. Prob level = probability level; PPV = positive predictive value; NPV = negative predictive value; F1 = harmonic mean of precision and recall.

Figure 1.

ROC curve for classification model to differentiate suicide from accidental intent among Maryland decedents* who died from overdose involving opioids between 2016 and 2020. ROC = receiver operating characteristic.

Comparisons of demographic and toxicology characteristics of reclassified suicide cases from the undetermined cohort when the classification model was applied with a probability cut-off of p(x) ≥ 0.5 and the original OCME designated suicide cases are presented in Table A2 in the Appendix. A total of 58 cases of the 7,772 undetermined cases were reclassified as suicide, representing a 43% increase in total suicide cases, but accounting for only 0.7% of total undetermined. Approximately 181 cases would be expected, assuming the ratio of suicide to accidental death was the same among undetermined decedents. As would be expected, however, the OCME classified suicide cases and the reclassified suicide cases did not differ significantly across the majority of characteristics.

Discussion

From a medical examiner perspective, manner of death determination is a highly complex and multidisciplinary process, requiring careful integration of known medical/psychological, investigative, scene, and toxicological information; because the outcome has lasting implications for families and the public record, it cannot be reduced to any single factor. Our cohort is small in comparison with the death toll of the opioid epidemic nationwide and only 1.4% of this sample of decedents was identified as suicide cases involving opioids. However, the demographics and toxicology findings among suicide decedents were substantially different, in aggregate, compared with either accidental or undetermined manner of death. A smaller proportion of suicide decedents were positive for fentanyl, cocaine, and morphine, but a larger share were positive for antidepressants, neuroleptics, oxycodone, benzodiazepines, and acetaminophen. The latter point illustrates a divergence in overdose manner of death partially through the presence of prescription medication. These findings should be interpreted in light of prior evidence that the lethality of drug classes varies significantly, with opioids and barbiturates demonstrating far higher fatality rates in overdose compared to substances such as antidepressants or NSAIDs (Miller et al., 2020). This suggests that observed MOD determination patterns may reflect both intent and pharmacologic lethality. For example, medical examiners may be more cautious in attributing suicide to highly lethal drugs that are also commonly used recreationally, where low doses can be fatal and circumstantial evidence may be sparse.

Notably, unlike a prior statewide analysis of suicide and undetermined deaths that found Black decedents were disproportionately classified as undetermined, we did not observe statistically significant racial differences in manner of death within our opioid-only cohort. This discrepancy may reflect differences in sample composition and data sources: our study was restricted to opioid-involved overdose deaths between 2016 and 2020, whereas Adams et al. (2025) examined all suicide and undetermined deaths from 2012 to 2020 and linked to health care utilization records.

The OCME findings enabled us to develop a fairly simple model of retrospective suicide risk for use in opioid-positive decedents, using easily attainable coded data without resource heavy reinvestigation, complex natural processing algorithms, or further data use agreements. High training precision is observed for a response p(x) > 0.5, and our sensitivity analysis illustrates sensitivities of greater than 50% of cases can be achieved by lowering the decision threshold. While the model successfully identified cases that could be reclassified as presumptive suicide among undetermined decedents, it fell short of our expected number of cases by roughly two-thirds. Our assumption was that there is an equal ratio of suicide cases to accidental among undetermined decedents as in the training data, however, the actual ratio could be much smaller. Although there is some expectation of bias among suicide cases to overuse of the undetermined MOD, it may be that this is just greater among accidental overdoses instead. It is impossible to know with certainty without a chart review and psychiatric autopsy of undetermined cases.

Because drug, especially opioid, concentrations vary with tolerance and postmortem changes, our model uses categorical (positive/high) indicators rather than raw levels, which cannot be interpreted in isolation. Additional clinical information might be used in this type of retrospective model through linkage of electronic health records (EHR), to substantially improve model performance. A greater precision and sensitivity could be achieved through highly associated diagnostic information, such as the presence of mood disorders, ADHD, and post-traumatic stress (Baldessarini & Tondo, 2020).

While these results are encouraging for the prospect of reducing false negative suicide classification, the demographics of Maryland may not be the most generalizable to conduct model development. The United States is large and diverse but Maryland is comparatively metropolitan, wealthy, educated, and ethnically diverse. This establishes a geographic context where social determinants of health work with health care availability to influence both the rate of overdose and suicide. It remains challenging to account for these effects with relatively small sample sizes, so additional modeling may be necessary at the national level, or locally with the linkage of vital statistics and EHR. While Maryland’s conservative classification practices limit the number of confirmed suicide training cases, they also yield one of the largest proportions of undetermined overdose deaths in the United States. This creates a uniquely valuable opportunity to identify potential suicides, yielding insights that would be difficult to obtain in jurisdictions where undetermined designations are rare. In this sense, Maryland offers a uniquely sensitive setting for detecting possible suicides that were classified otherwise, even if the findings require validation elsewhere.

Furthermore, our team is currently expanding this work by linking death records with electronic health records and conducting psychological autopsies for a subset of decedents in Maryland. This additional data will allow for stronger validation of the model and broader testing of lower probability thresholds, setting the stage for more comprehensive suicide surveillance.

Limitations

Finally, the lack of truth labels for undetermined cases is very limiting in this analysis. Without independent confirmation of intent, the accuracy of reclassification cannot be definitively established. The necessary next step to this effort would be to perform psychological autopsies and chart review of clinical records to assess whether corroborating information can be found for reclassified cases. Psychological autopsy, in which semistructured interviews with next of kin and other collateral informants are used to clarify the proximal circumstances and mental state of the deceased, has been cited by both the CDC and the American Association of Suicidology as the best practice for verifying most likely manner of death and can be used in future studies to validate the reclassification of cases (American Association of Suicidology, 2024; Cavanagh et al., 2003; Stone et al., 2017). Incorporating such investigations, alongside chart reviews, will be essential in future work to confirm reclassified cases and refine model performance for broader application.

Conclusion

Medical examiners and coroners often struggle to determine the likelihood of suicide among overdose decedents. This problem is particularly pronounced in Maryland, which holds the highest proportion of undetermined manner overdose deaths in the United States. Since deaths classified as undetermined are not actionable for policy and programming due to their heterogeneity, this creates a critical gap in public health planning. Leveraging basic demographic and toxicology data from standard autopsies only, we developed a model that detected 43% more suicide deaths with good statistical precision. As a next step, psychological autopsy could be used to confirm MOD for these decedents and to test the potential use of this model at lower probability thresholds for more comprehensive detection of missed suicide deaths.

This model does not supplant medicolegal judgment but may serve as a screening tool to identify patterns and cases meriting further review. Its primary utility lies in enhancing surveillance and public health understanding of suicide among opioid deaths, thereby informing prevention and resource allocation. While our model provides a valuable starting point for detecting unclassified suicides among overdose deaths, its limitations are clear. Incorporating clinical information, such as mental health history or evidence of prior suicidal ideation, could enhance predictive accuracy and improve classification efforts. Future work should explore these avenues, including the potential for using linked electronic health records or psychological autopsies to validate and refine our approach. Ultimately, a more comprehensive data set will further address the significant underreporting of suicide deaths in the context of opioid overdoses, both in Maryland and nationally.

Biographies

Jean Flores, DrPH, is an experienced researcher in youth suicide prevention and has served as measurement strategy lead for Kaiser Permanente’s national suicide prevention initiative and pediatric firearm safety initiative. She currently works as an epidemiology and biostatistics methods consultant at Kaiser Permanente Northern California’s Division of Research.

Monica M. Desjardins, MPH (Diné), is an epidemiologist at RTI International and a doctoral student at Johns Hopkins Bloomberg School of Public Health. She has five years of experience in suicide and substance use prevention and is committed to reducing mental health disparities in Indigenous communities through research and advocacy.

Christopher Kitchen, MS, is a senior research analyst for the Center for Population Health IT at the Johns Hopkins Bloomberg School of Public Health. He specializes in machine learning and has 10 years of experience in the psychiatric assessment of patients with substance use disorders and schizophrenia-related diseases.

Anas Belouali, MEng, MS, is a health data scientist and an NLM T15 fellow pursuing a PhD in biomedical informatics and data science at Johns Hopkins School of Medicine. His research focuses on suicide risk prediction and subtyping using health records, with over a decade of experience across healthcare, academia, and industry.

Hadi Kharrazi, MD, PhD, is the co-director of the Johns Hopkins Center for Population Health IT (CPHIT), which focuses on advancing the use of IT in various areas of population health. His research specifically focuses on the application of informatics solutions to advance the science of population health analytics.

Holly Wilcox, PhD, is the director and founder of the Center for Suicide Prevention at the Johns Hopkins Bloomberg School of Public Health. Her research focuses on advancing public health approaches to suicide prevention, including policies, early intervention, and chain of care approaches.

Paul Nestadt, MD, is the James Wah Professor of Psychiatry at Johns Hopkins and medical director of its Center for Suicide Prevention. A nationally recognized expert on suicide risk and prevention, his research focuses on firearms, opioids, and psychological autopsy, while his clinical work spans anxiety and treatment-resistant illness.

Appendix

Table A1.

Demographic and toxicology characteristics of Maryland opioid overdose decedents between 2016 and 2020

	Overall		Suicide		Accident		Undetermined		p valuea Accident versus undetermined	p value All manner of death (MOD)
	N = 10,218		N = 134		N = 2,312		N = 7,772
	100%		1.4%		22.5%		76.1%
Characteristics	n	%	n	%	n	%	n	%	(L–R χ2)	(L–R χ2)
Male	7,442	73	60	45	1,699	73	5,683	73	.75	<.0001
Age category									.02	<.0001
<24 years	632	6	10	7	117	5	505	7
25–39 years	3,784	37	25	19	838	36	2,921	38
40–64 years	5,387	53	64	48	1,264	55	4,059	52
65+	415	4	35	26	93	4	287	4
Race									.53	<.0001
White	6,298	62	116	87	1,419	61	4,763	61
Black	3,551	35	13	10	801	35	2,737	35
Other	369	4	5	4	92	4	272	4
NCHS									.22	<.0001
Large central metro	2,695	26	17	13	632	27	1,119	14
Large fringe metro	5,056	49	97	72	1,156	50	2,046	26
Medium metro	561	5	7	5	109	5	3,803	49
Small metro	296	3	4	3	69	3	445	6
Micropolitan	84	1	2	1	21	1	223	3
Noncore	92	1	2	1	15	1	61	1
Missing	1,434	14	5	4	310	13	75	1
Class: Fentanyl positiveb	8,889	87	28	21	1,912	83	6,949	89	<.0001	<.0001
Class subset: Fentanyl positivec	8,343	82	27	20	1,806	78	6,510	84	<.0001	<.0001
Class subset: Fentanyl highd	381	4	1	1	80	3	300	4	.37	.06
Class: Cocaine positive	3,946	39	12	9	870	38	3,064	39	.12	<.0001
Class subset: Cocaine positive	3,484	34	10	7	751	32	2,723	35	.02	<.0001
Class subset: Cocaine high	539	5	3	2	116	5	420	5	.46	.16
Class: Oxycodone positive	1,246	12	84	63	347	15	815	10	<.0001	<.0001
Class subset: Oxycodone positive	1,246	12	84	63	347	15	815	10	<.0001	<.0001
Class subset: Oxycodone high	80	1	34	25	14	1	32	0.4	.24	<.0001
Substance: Oxycodone positive	1,024	10	65	49	285	12	674	9	<.0001	<.0001
Substance: Oxycodone high	55	0.5	25	19	7	0.3	23	0.3	.95	<.0001
Substance: Oxymorphone positive	127	1	9	7	43	2	75	1	.0009	<.0001
Substance: Oxymorphone high	7	0.1	3	2	3	0	1	0.01	.003	<.0001
Substance: Hydrocodone positive	141	1	22	16	31	1	88	1	.42	<.0001
Substance: Hydrocodone high	19	0.2	11	8	2	0	6	0	.89	<.0001
Substance: Hydromorphone positive	81	0.01	12	9	24	1	45	1	.03	<.0001
Substance: Hydromorphone high	5	0.1	1	1	2	0.1	2	0.03	.24	.08
Class: Heroin/Morphine positive	4,466	44	49	37	950	41	3,467	45	.003	.003
Class subset: Heroin/Morphine positive	3,592	35	44	33	748	32	2,800	36	.001	.004
Class subset: Heroin/Morphine high	18	0.2	2	1	3	0.1	13	0.2	.68	.07
Class/Substance: Acetaminophen positive	665	7	40	30	170	7	455	6	.01	<.0001
Class/Substance: Acetaminophen high	24	0.2	12	9	3	0.1	9	0.1	.87	<.0001
Class: Neuroleptics positive	598	6	17	13	146	6	435	6	.20	.001
Class: Benzodiazepines positive	1,447	14	39	29	359	16	1,049	14	.01	<.0001
Class: Methadone positive	876	9	4	3	223	10	649	8	.05	.005
Class: Ethanol positive	3,469	34	44	33	728	31	2,697	35	.004	.0154
Class: Antidepressants positive	2,080	20	67	50	483	21	1,530	20	.21	<.0001
Substance: Amitriptyline positive	180	2	7	5	34	1	139	2	.29	.0255
Substance: Amitriptyline high	13	0.1	3	2	2	0.1	8	0.1	.82	.0021
Substance: Bupropion positive	257	3	14	10	64	3	179	2	.21	<.0001
Substance: Bupropion high	37	0.4	8	6	6	0.3	23	0.3	.77	<.0001
Substance: Citalopram positive	592	6	11	8	152	7	429	6	.06	.0889
Substance: Citalopram high	21	0.2	3	2	2	0.1	16	0.2	.20	.0041
Substance: Doxepin positive	91	1	2	1	21	1	68	1	.88	.7839
Substance: Doxepin high	10	0.1	2	1	2	0.1	6	0.1	.89	.0225
Substance: Fluoxetine positive	311	3	10	7	57	2	244	3	.09	.0091
Substance: Fluoxetine high	26	0.3	0	0	5	0.2	21	0.3	.65	.6385
Substance: Mirtazapine positive	230	2	12	9	66	3	152	2	.01	<.0001
Substance: Mirtazapine high	11	0.1	3	2	4	0.2	4	0.1	.10	.0003
Substance: Paroxetine positive	85	1	2	1	22	1	61	1	.44	.5573
Substance: Paroxetine high	4	0	0	0	2	0.1	2	0	.24	.4738
Substance: Sertraline positive	370	4	13	10	72	3	285	4	.20	.0029
Substance: Sertraline high	32	0.3	1	1	4	0.2	27	0.4	.16	.2727
Substance: Trazodone positive	606	7	9	7	147	6	450	6	.31	.5596
Substance: Trazodone high	39	0.4	6	4	3	0.1	30	0.4	.04	<.0001
Substance: Venlafaxine positive	157	2	10	7	38	2	109	1	.40	.0002
Substance: Venlafaxine high	18	0.2	2	1	3	a.1	13	0.2	.68	.0675

Note.

^aLikelihood ratio chi square or Fisher exact test comparison of subgroups.

^bClass = At least one substance in class is positive.

^cClass subset: Positive = among substances with continuous results in class, at least one substance is positive.

^dClass subset: High = among substances with continuous results in class, at least one substance meets or exceeds “high” toxicity threshold.

Table A2.

Comparison of demographic and toxicology characteristics of Maryland opioid overdose decedents between 2016 and 2020 between (1) OCME confirmed suicide decedents and (2) OCME undetermined cases reclassified as suicide decedents using predictive model with probability cut-off of p(x) ≥ 0.5

	OCME suicide		Reclassified suicide
	N = 134		N = 58
Characteristics	n	%	n	%	p valuea
Male	60	44	17	29	.05
Age category					.76
<24 years	10	7	5	9
25–39 years	25	19	14	24
40–64 years	64	48	27	47
65+	35	26	12	21
Race					.28
Black	13	10	10	17
Other	5	4	1	2
White	116	87	47	81
NCHS					.11
Large central metro	17	13	5	9
Large fringe metro	97	72	37	64
Medium metro	7	5	5	9
Small metro	4	3	1	2
Micropolitan	2	1	1	2
Noncore	2	1	0	0
Missing	5	4	9	16
Class: Fentanyl positiveb	28	21	9	16	.43
Class subset: Fentanyl positivec	27	20	8	14	.41
Class subset: Fentanyl highd	1	1	0	0	>.99
Class: Cocaine positive	12	9	4	7	.78
Class subset: Cocaine positive	10	7	3	5	.76
Class subset: Cocaine high	3	2	1	2	>.99
Class: Oxycodone positive	84	63	44	76	.10
Class subset: Oxycodone positive	84	63	44	76	.10
Class subset: Oxycodone high	34	25	25	43	.02
Substance: Oxycodone positive	65	49	35	60	.16
Substance: Oxycodone high	25	19	20	34	.03
Substance: Oxymorphone positive	9	7	3	5	>.99
Substance: Oxymorphone high	3	2	0	0	.55
Substance: Hydrocodone positive	22	16	7	12	.52
Substance: Hydrocodone high	11	8	5	9	>.99
Substance: Hydromorphone positive	12	9	3	5	.56
Substance: Hydromorphone high	1	1	0	0	>.99
Class: Morphine positive	49	37	14	24	.10
Class Subset: Morphine positive	44	33	13	22	.17
Class Subset: Morphine high	2	1	0	0	>.99
Class/Substance: Acetaminophen positive	40	30	18	31	.87
Class/Substance: Acetaminophen high	12	9	6	10	.79
Class: Neuroleptics positive	17	13	12	21	.19
Class: Benzodiazepines positive	39	29	22	38	.24
Class: Methadone positive	4	3	1	2	>.99
Class: Ethanol positive	44	33	16	28	.50
Class: Antidepressants positive	67	50	46	79	<.001
Substance: Amitriptyline positive	7	5	3	5	>.99
Substance: Amitriptyline high	3	2	0	0	.55
Substance: Bupropion positive	14	10	16	28	<.01
Substance: Bupropion high	8	6	11	19	<.01
Substance: Citalopram positive	11	8	6	10	.56
Substance: Citalopram high	3	2	1	2	>.99
Substance: Doxepin positive	2	1	4	7	.07
Substance: Doxepin high	2	1	0	0	>.99
Substance: Fluoxetine positive	10	7	9	16	.11
Substance: Fluoxetine high	0	0	1	2	.30
Substance: Mirtazapine positive	12	9	8	14	.31
Substance: Mirtazapine high	3	2	1	2	>.99
Substance: Paroxetine positive	2	1	1	2	>.99
Substance: Paroxetine high	0	0	0	0	>.99
Substance: Sertraline positive	13	10	6	10	>.99
Substance: Sertraline high	1	1	0	0	>.99
Substance: Trazodone positive	9	7	17	29	<.001
Substance: Trazodone high	6	4	15	26	<.001
Substance: Venlafaxine positive	10	7	5	9	.78
Substance: Venlafaxine high	2	1	0	0	>.99

Note.

^aLikelihood ratio chi square or Fisher exact test comparison of subgroups.

^bClass = At least one substance in class is positive.

^cClass subset: Positive = among substances with continuous results in class, at least one substance is positive.

^dClass subset: High = among substances with continuous results in class, at least one substance meets or exceeds “high” toxicity threshold.

NCHS = National Center for Health Statistics. OCME = Office of the Chief Medical Examiner of Maryland.