Prediction of adverse events risk in patients with comorbid post-traumatic stress disorder and alcohol use disorder using electronic medical records by deep learning models

Abstract

Background:

Identifying co-occurring mental disorders and elevated risk is vital for optimization of healthcare processes. In this study, we will use DeepBiomarker2, an updated version of our deep learning model to predict the adverse events among patients with comorbid post-traumatic stress disorder (PTSD) and alcohol use disorder (AUD), a high-risk population.

Methods:

We analyzed electronic medical records of 5565 patients from University of Pittsburgh Medical Center to predict adverse events (opioid use disorder, suicide related events, depression, and death) within 3 months at any encounter after the diagnosis of PTSD+AUD by using DeepBiomarker2. We integrated multimodal information including: lab tests, medications, co-morbidities, individual and neighborhood level social determinants of health (SDoH), psychotherapy and veteran data.

Results:

DeepBiomarker2 achieved an area under the receiver operator curve (AUROC) of 0.94 on the prediction of adverse events among those PTSD+AUD patients. Medications such as vilazodone, dronabinol, tenofovir, suvorexant, modafinil, and lamivudine showed potential for risk reduction. SDoH parameters such as cognitive behavioral therapy and trauma focused psychotherapy lowered risk while active veteran status, income segregation, limited access to parks and greenery, low Gini index, limited English-speaking capacity, and younger patients increased risk.

Conclusions:

Our improved version of DeepBiomarker2 demonstrated its capability of predicting multiple adverse event risk with high accuracy and identifying potential risk and beneficial factors.

1. Background

According to the Substance Abuse And Mental Health Administration, 20.8 million Americans have substance use disorder (SUD) of which 15.7 million experience alcohol use disorder (AUD), making alcohol the most commonly abused substance in the United States (Nehring, 2022). Almost half of these patients experienced co-existing mental illnesses including posttraumatic stress disorder (PTSD). PTSD patients are two or three times more likely to experience substance use disorder (Berenz, 2012). Compared to AUD or PTSD alone, patients with co-occurring PTSD and AUD exhibit greater symptom severity, poorer quality of life, lower recruitment and non-compliance in treatment programs, poorer clinical outcomes, and faster relapses in post-treatment surveillance programs (Coffey et al., 2016). AUD is a common mental health condition with increased relapse tendencies in PTSD patients. It impacts (1) positive effect regulation (e.g., increased alcohol intake perpetuating feelings of euphoria), (2) negative effect regulation (e.g., coping with negative feelings such as anxiety, depression, feeling of worthlessness may increase loss of control over alcohol intake), (3) pharmacological vulnerability (e.g., the body’s ability to cope with alcohol metabolism) and (4) deviance proneness (e.g., an individual’s deviant behavior due to social history, childhood trauma etc.)(Nehring, 2022). -Co-morbid alcohol use disorder (AUD) and PTSD are frequently linked to negative outcomes or indicators of poor outcomes. For example: patients with co-occurring AUD and PTSD were more likely to be females, sexual minorities, veterans, and face socio-economic issues including homelessness, unemployment, poverty, lack of access to education, health care, and food with a history of incarceration and sexual abuse (Bryant-Davis, 2010; Smith, 2018). A recent study found an increased likelihood of experiencing other psychiatric co-morbidities coupled with worse social and psychiatric functioning in patients with both PTSD and AUD (16.1% of the population) as opposed to PTSD alone (8.5% of the population) (Simpson, 2019). Another recent report by the U.S. Veteran Affairs (VA) healthcare system found veterans with PTSD+AUD had increased risk of suicide as opposed to their PTSD alone counterparts suggesting the need to improve early diagnoses and interventions for these patients (Blakey, 2022).

Identification of medications targeting both PTSD and AUD is a key area in the field. Current efforts include: (1) Using FDA-approved pharmacological therapies for either disorders alone: Naltrexone, Sertraline (2) Pharmacological therapies targeting the underlying pathophysiology shared by both conditions: Prazosin, Topiramate, Desipramine, Zonisamide, (3) Behavioral approaches for PTSD and AUD: Eye Movement Desensitization and Reprocessing (EMDR) therapy, seeking safety, cognitive behavioral therapy, exposure therapy, and other interventions (4) Alternate approaches: Mindfulness, acupuncture, motivational enhancement therapy, and yoga. Both PTSD and AUD alter monoamine neurotransmission pathways impacting dopamine, norepinephrine, and serotonin levels within the hypothalamic-pituitary adrenal axis and the meso-cortico-limbic systems (Verplaetse, 2018). These neural circuits are heavily involved in the regulation of stress and addiction. This evidence reiterates the clinical need to search for other treatment options that combine pharmacological, behavioral, and alternative interventions to optimize treatment efficacy in these patients. Unsurprisingly, there are no FDA-approved treatments for co-occurring PTSD and AUD. While various studies suggest that both PTSD and AUD share common dysfunctions, the clinical physiology of patients with both conditions mimic patients with PTSD more than AUD. This suggests PTSD plays a greater role in increased social, emotional, and neuropsychiatric impairment which in turn enhances AUD vulnerability. There is a lack of evidence on personalized treatment for co-morbid PTSD and AUD, due to the exclusion of these high-risk patients in clinical trials. These patients experience persistent and severe risk of adverse events (e.g., treatment resistance symptoms, advancement to other SUDs, suicide, and death), thus emphasizing the need to improve treatment options suitably tailored for all (Ralevski, 2014). This can be achieved via application of novel analytic technologies to data-mine electronic medical record (EMR) data from PTSD+AUD patients.

Both PTSD and AUD have detrimental impact on various aspects of life. While reducing alcohol consumption and improving overall health can lead to recovery from AUD, it is also essential to address the associated social and psychiatric challenges through pharmacological interventions and alternative approaches. Social determinants of health (SDoH) are “conditions or environments in which people are born, grow, live, work, and age,” (Korn et al. 2022). Five key SDoH domains that have significant impacts on human health are (1) economic stability, (2) education, (3) health and health care, (4) neighborhood and built environment, and (5) social and community context (Smiley Y et al. 2022). Previous studies have focused on individual-level SDoH parameters and found improvement in financial status, employment status, access to healthcare, and community support systems have an increased likelihood of recovery. Other community-level SDoH factors found communities with lower socioeconomic status had higher access to alcohol that was associated with greater alcohol consumption and exacerbated other harmful behaviors. However, to date, little research has examined the association of AUD- and PTSD-related problems at less granular levels of SDoH influence (e.g., communities) on adverse events. There is a large research and clinical gap in the co-occurring PTSD and AUD field. Lack of (1) Dissemination of multi-modal factors that contribute to increased/decreased risk of adverse events in diverse populations (e.g., civilian, veterans, underserved, pregnant) (2) Identification of context-specific SDoH parameters (e.g. policies pertaining to career consequences such as recording clinical trial/treatment participation that may lead to employment issues) (3) Comparative efficacy studies for identification of treatments for targeted populations, (4) Inclusion of high risk population in prospective prevention studies due to difficulty in recruitment and retention and (5) Drug repurposing studies using SUD treatments to increase treatment engagement and prevent treatment dropout. We aim to explore which treatments work best for which individual, and if matching each treatment option to patient characteristics improves overall outcomes.

EMRs are crucial for clinical practice and documentation, but they provide limited SDoH information, which strongly impacts mental health. Few studies have explored the predictive value of the multimodal information extracted from EMRs, such as diagnoses, medication use, laboratory test results, individual-level SDoH indicators (e.g., race, age, gender, etc.), and community/neighborhood-level SDoH indicators (e.g., nSES index, etc.) for high-risk patient outcomes (Fan, 2023; Miranda, 2022a,b, 2023). Our overarching goal is to leverage EMRs to predict risks and develop evidence-based interventions that allow healthcare practitioners to customize treatment plans for each patient. These interventions, based on biomarkers, demographics, co-morbidities, medication use, psychotherapy status, veteran status, and other characteristics, aim to identify alternative options if conventional treatments are ineffective.

Deep learning/data mining algorithms are powerful tools for extracting valuable insights from large-scale EMR data (Wang and Raj, 2017). These algorithms can uncover hidden patterns and dependencies in the data, enabling the automatic discovery of important relationships that may have otherwise been overlooked (Shickel et al., 2018). By applying these techniques to analyze the EMR of PTSD patients, we can identify risk factors and medications that have the potential to significantly impact disease progression. In this study, we have applied our current model DeepBiomarker2 to predict adverse event outcomes related to PTSD+AUD. The refined results obtained from DeepBiomarker2 are specific to high-risk cohorts, providing valuable interdisciplinary hypotheses and identifying medications that can potentially prevent adverse events.

2. Methods

2.1. Data source

We extracted data from the Neptune system at the University of Pittsburgh Medical Center (UPMC) covering the period from January 2004 to October 2020 (rio.pitt.edu/services). The database contains a rich set of multimodal information, including demographic details, diagnoses, medication prescriptions, and laboratory test results. In our study, the patients were identified using specific ICD9/10 codes (refer to Appendix A for details)(Heslin KC and Steiner 2006).

2.2. Data preparation

The data preparation process followed a similar approach as described in our previous publication, DeepBiomarker and DeepBiomarker2 (Miranda, 2022a,b; Miranda, 2023). We aimed to predict new diagnoses of adverse events in a 3-month period at any encounter (index date) for patients who already has PTSD+AUD diagnosis (Fig. 2). Cases had adverse events within 3 months of index date, while controls had no adverse event records during the same period. We applied data augmentation in the data preparation stage. For cases, there is a possibility that the patient had multiple encounters. We included all the eligible encounters that occurred between the first diagnoses of PTSD+AUD and the new onsets of adverse events; for controls, we randomly picked nearly the same number of samples from encounters after the first diagnoses of PTSD+AUD. To include more patients, we also included as cases patients who already had one or more of the adverse events before or at the initial diagnoses of PTSD+AUD diagnoses but developed new other types of adverse events of interest after the index date. We used multimodal information, including diagnoses, medications, lab test results, SDoH, psychotherapy and veteran data. Lab test results categorized as “ABNORMAL”, “HIGH”, or “LOW” were included, excluding low-frequency tests. Diagnoses were clustered, and medication names were transformed into unique DrugBank IDs. Multimodal information was organized into sequences based on disease categories, DrugBank IDs, and lab test IDs.

Fig. 2.

Overview of DeepBiomarker2. (A) Data sampling from electronic medical records: Patient A and B both pass the inclusion criteria and within the given time interval, Patient A has no event and Patient B has any one of the adverse events and are considered as a control and a case respectively. We extract their multimodal information (i.e. Diagnoses/Disease, Medication use/Drug and Lab-test results from their structured EMRs and use them as input in our model, (B) Data embedding: The multimodal information is then converted into continuous vector spaces to build an embedding matrix and (C) Prediction by neural networks that include TLSTM and RETAIN as the basic prediction units: We incorporated individual and neighborhood level SDoH information, trauma focused psychotherapy, cognitive behavioral status and veteran status information for outcome prediction. Our model provides us with a comprehensive list of multiple biomarkers. We employ perturbation-based contribution analysis technique to calculate the relative contribution value and identify the most important features/biomarkers depending upon their relative contribution (RC) value. If RC>1, then that biomarker is an indicator of high risk, RC<1, then that biomarker is an indicator of low risk. LSTM: Long Short-Term Memory; TLSTM: Time aware long short-term memory; RETAIN: Reverse Time Attention Model; Dx: diagnosis.

2.3. Dataset splitting

We partitioned our dataset into random training, validation, and test subsets using an 8:1:1 ratio.

2.4. SDoH data

We incorporated both individual-level and neighborhood-level SDoH data for each PTSD+AUD patient. Individual-level SDoH features such as race, age, and gender were extracted from the demographic information in the structured EMR data and encoded similarly to diagnoses codes for model input. “Neighborhood-level SDoH features (included in supplementary information) were calculated using specific formulas and 5-year estimates data from the American Community Survey (ACS) and other datasets. We used patients’ 5-digit ZIP codes from the EMR data at the index date to map their neighborhood-level SDoH parameters.” The data were aggregated and utilized as input for our model, aligning with previous studies that have employed similar indexes to assess the socioeconomic deprivation of geographical areas (CA., 2002) (Bhavsar NA et al. 2018).

2.5. Psychotherapy

Cognitive behavioral therapy (CBT) and trauma-focused psychotherapy data were extracted for each PTSD patient from unstructured EMR data. Trauma-focused psychotherapy encompassed cognitive processing therapy (CPT), prolonged exposure (PE), and eye movement desensitization and reprocessing (EMDR). We employed the same extraction process that involved creating a custom keyword dictionary (Fig. 1), (Miranda, 2023) identifying relevant sentences containing specific keywords related to these therapies, developed our custom sentence dictionary and used a pre-trained sentence transformer model (all-mpnet-base-v2) for further identification of patients with/without psychotherapy(“Sentence-Transformers, 2021). Instances of CBT and trauma-focused psychotherapy were identified, coded, and used as SDoH input in our models for patients already identified as cases and controls.

Fig. 1.

Iterative workflow of extraction of psychotherapy and veteran information from clinical notes (A) Data sampling and preprocessing: We leveraged 5.7 million clinical notes from UPMC EMR system to extract multimodal information (i.e. psychotherapy status and veteran information to be used as input in our model, (B) Building of custom keyword dictionary: We created a custom keyword dictionary with the help of literature and subject matter experts (C) Building of custom sentence dictionary: Next, we extracted sentences around those keywords and with the help of subject matter experts developed our own custom sentence dictionary. (D) Identification of patients with psychotherapy and veteran information using sentence transformers: We calculated the cosine similarity scores between our custom sentence dictionary (U) and sentences from clinical notes (V), to identify which patients underwent or are undergoing psychotherapy and are veterans. This information was then binary coded and used as input in our model. *UPMC: University of Pittsburgh Medical Center, EMR: electronic medical records, BERT: Bidirectional encoder representations from transformers, SMEs: Subject Matter Experts.

2.6. Veteran status

We also identified veterans from the PTSD clinical notes using the same natural language processing technique that we employed to identify psychotherapy status. We created a custom sentence dictionary and included only those sentences that have the following keywords: “veteran”, “military”, “army”, “navy”, “air-force”, “service member”, “Veterans affairs”, “marines” and “coast guard”. We followed the same process as mentioned in the psychotherapy section and identified them and then used them as SDoH input in our models.

Iterative workflow of extraction of psychotherapy and veteran information from clinical notes.

2.7. DeepBiomarker2

We utilized the Pytorch_EHR framework developed by the Zhi Group (Chang et al., 2017), which employs multiple recurrent neural networks such as time aware LSTM and RETAIN for the analysis and prediction of clinical outcomes (Rasmy et al., 2021). LSTM is a recurrent neural network variant designed for capturing long-range dependencies in sequential data. It is comprised of input, forget, and output gates, which control information flow. The default activation function is sigmoid, limiting outputs to a range of 0–1(add paper). RETAIN analyses EMR data and incorporates an attention mechanism to focus on important events in a patient’s medical history. The model captures temporal dependencies and assigns weights to relevant events (Mechanism, 2016). Building upon this framework and our previous DeepBiomarker models, DeepBiomarker and DeepBiomarker2, we applied it to the current dataset (Miranda, 2022a,b; Miranda, 2023). We integrated individual lab test results, SDoH parameters, psychotherapy data, veteran data, medications, and diagnoses as input just as mentioned in our previous versions. Additionally, we calculated the relative contribution of identified factors by evaluating the observed changes in the model resulting from random perturbations in those factors (Guan et al., 2019). For consistency with our previous versions, we maintained the same parameter settings: embedding dimensions of 128, hidden size of 128, dropout rate of 0.2, number of layers set to 8, and a patience value of 3. We calculated the standard deviation of our accuracy from ten cross-validated iterations of our algorithm.

3. Statistical analysis

3.1. Assessment of importance of the multiple clinical factors for predicting adverse events

To examine the importance of the clinical factors for prediction of adverse events, we calculated the relative contribution (RC) of each feature on the adverse events (Guan et al., 2019). The RC of a feature was calculated as the median contribution of the feature to events divided by the median contributions of this feature to no-events. The contributions were estimated by a perturbation-based approach (Rao et al., 2020). The RC value and significance calculation are shown as follows where FC represents the feature contribution:

$$\begin{gathered} \text{RC}\text{value}=\frac{\text{median}\left({\text{FC}}_{\text{with}\text{event}}\right)}{\text{median}\left({\text{FC}}_{\text{without}\text{event}}\right)} \\ \text{RC}\text{significance}= \\ \text{Wilcoxon}\text{rank}\text{sum}\text{test}\text{p}\text{value}\left({\text{FCs}}_{\text{with}\text{event}}{\text{andFCs}}_{\text{without}\text{event}}\right) \end{gathered}$$

FC value was the total value of the feature within the same patient if the feature appeared more than once in that patient. In most cases, the FCs were not normally distributed. As such, FC medians were used in the RC calculation instead of FC means and the significance of RCs were represented by the p-value of a Wilcoxon rank sum test comparing median FCs with and without events (Wilcoxon, 1945). The false discovery rate (FDR) was used to reduce the type I error caused by multiple comparisons. FDR is the expected ratio of the number of false positive results to the number of total positive test results (Pounds S, 2004). An FDR of 0.05 was used to determine significance.

We enhanced our assessment by normalizing the FC value and scaling the RC value for all our features. The improved FC formula is calculated as the ratio of the summary of the contribution of a feature and the summary of the contribution of all the features. For, the RC of a PTSD diagnosis, we scaled it to 1 and was applied to get the final RC value for each of the other features. FC normalization accounts for variations in the numbers of encounters across patients, preventing the overemphasis on the contribution of features observed in patients with frequent healthcare visits. Scaling helps to differentiate between indicators of high risk from indicators of low risk by using a reference point (PTSD diagnosis). Since raw RCs can be influenced by the imbalance in visits among cases and controls, this approach provides a fair assessment. It is essential to note that this field of model interpretation is evolving, and more complex methods may emerge. To validate our approach, we can leverage prior knowledge to assess the performance of analysis, discovering unknown patterns and generating new hypotheses.

3.2. Assessment of model performance

The model performance was evaluated by the area under the ROC curve (AUROC).

4. Results

4.1. The performance of DeepBiomarker2 on the adverse events prediction

We utilized UPMC EMR data to identify a total of 5565 PTSD+AUD patients from 38,807 PTSD patients. From this cohort, we further identified 7927 cases and 7685 controls. We then assessed the performance of DeepBiomarker2 which can be found in Table 1.

Table 1.

The performance of DeepBiomarker2 with and without SDoH features.

RETAIN (+SDOH)	1	2	3	4	5	average	std.s
Validation AUC	0.95	0.96	0.96	0.96	0.96	0.96	0.018
Test AUC	0.95	0.96	0.95	0.95	0.95	0.95	0.003
Test Precision	0.90	0.92	0.94	0.91	0.90	0.90	0.015
Test Recall	0.90	0.90	0.86	0.89	0.90	0.90	0.017
Test F1	0.90	0.91	0.89	0.90	0.90	0.90	0.006
RETAIN(−SDOH)	1	2	3	4	5	average	std.s
Validation AUC	0.95	0.96	0.96	0.96	0.97	0.96	0.006
Test AUC	0.95	0.95	0.95	0.95	0.95	0.95	0.001
Test Precision	0.87	0.89	0.90	0.89	0.90	0.89	0.015
Test Recall	0.92	0.91	0.90	0.89	0.89	0.90	0.013
Test F1	0.89	0.90	0.90	0.89	0.90	0.90	0.003
TLSTM(+SDOH)	1	2	3	4	5	average	std.s
Validation AUC	0.97	0.97	0.97	0.97	0.96	0.97	0.002
Test AUC	0.95	0.95	0.95	0.96	0.95	0.95	0.003
Test Precision	0.89	0.88	0.86	0.87	0.87	0.87	0.013
Test Recall	0.90	0.90	0.93	0.93	0.91	0.91	0.016
Test F1	0.89	0.89	0.89	0.90	0.89	0.89	0.004
TLSTM(−SDOH)	1	2	3	4	5	average	std.s
Validation AUC	0.96	0.97	0.97	0.97	0.96	0.97	0.004
Test AUC	0.95	0.95	0.96	0.96	0.95	0.95	0.005
Test Precision	0.83	0.85	0.89	0.93	0.90	0.88	0.037
Test Recall	0.92	0.90	0.90	0.89	0.89	0.90	0.014
Test F1	0.88	0.88	0.89	0.91	0.89	0.89	0.013

^*AUC: area under the curve; sd: standard deviation, T-LSTM: Time aware Long Short-Term Memory; RETAIN: Reverse Time AttentIoN model; +SDOH/−SDOH includes/excludes social determinants of health factors.

Table 1 presents the performance results of the DeepBiomarker2 model, which incorporates deep learning algorithms such as LSTM and RETAIN, for predicting adverse events. The deep learning models demonstrate excellent performance, as indicated by AUC scores ≥0.90. These findings reinforce the importance of leveraging advanced deep learning techniques and considering multimodal information in predictive modeling for improved clinical outcomes.

4.2. Important biomarkers for adverse events prediction

In our analysis, we utilized a perturbation-based estimation approach to determine the relative contribution of input features in predicting adverse events (Supplementary Table 1). The results of this analysis are presented in Table 2, Table 3, Table 4, and Table 5, which highlight the top important abnormal lab tests, medication use, diagnoses, and SDoH parameters, respectively. Table 2 demonstrates that all abnormal lab test results exhibit an RC>1, indicating that they are classified as risk factors for adverse events. As shown in Table 3, medications such as omeprazole, gabapentin, albuterol, ibuprofen hydrocodone, and oxycodone have an RC>1, signifying their association with increased risk for adverse events. Other hand, medications such as piroxicam, vilazodone, dronabinol, tenofovir, suvorexant, empagliflozin, famciclovir, veramyst, amantadine, sulfasalazine, and lamivudine have an RC<1, indicating that they are categorized as indicators of low risk. In Table 4, diagnoses pertaining to pain, neuroinflammation, and home accidents were indicators of high risk, while other counseling was found to be indicator of low risk. In Table 5, patients who had undergone/undergoing psychotherapy had lower risk of adverse events, while patients who were identified as veterans had higher risk of adverse events. These findings provide valuable insights into the potential impact of biomarkers on the prediction of adverse events in PTSD+AUD patients. The identification of risk and protective factors can inform healthcare professionals in designing more targeted interventions and treatment plans to mitigate the occurrence of adverse events among PTSD+AUD patients.

Table 2.

Significant abnormal lab test results identified through perturbation-based contribution analysis for predicting adverse events.

Name	RC	Wilcoxon_p	FDR_Q
Glucose	1.48	9.27E-126	9.27E-126
Albumin	1.70	1.30E-92	3.24E-93
Red blood cells	1.46	1.29E-85	2.58E-86
Hematocrit	1.43	9.83E-74	1.43E-74
Hemoglobin	1.41	1.00E-73	1.43E-74
Calcium	1.60	5.94E-70	7.42E-71
Chloride	1.48	5.27E-65	5.86E-66
AST	1.53	1.97E-61	1.97E-62
Sodium	1.50	1.28E-58	1.07E-59
Carbon dioxide	1.64	3.56E-56	2.74E-57
Creatinine	1.64	2.41E-54	1.72E-55
Potassium	1.48	8.52E-53	5.68E-54
MCHC	1.47	6.64E-52	3.91E-53
RDW	1.42	4.63E-51	2.57E-52
Urea nitrogen	1.48	1.17E-50	6.16E-52
Platelets	1.56	4.89E-50	2.45E-51
MCH	1.45	1.20E-46	5.72E-48

^*red cell distribution width (RDW), aspartate aminotransferase (AST), mean corpuscular hemoglobin concentration (MCHC), and mean corpuscular hemoglobin (MCH). Relative contribution value > 1: Risk and Relative contribution value < 1: Protective; FDR_Q: false discovery rate adjusted Q value; Wilcoxon_p: P values with Wilcoxon.

Table 3.

Significant medications identified through perturbation-based contribution analysis for predicting adverse events.

Name	RC	Wilcoxon_p	FDR_Q
Hydrocodone	1.48	3.40E-23	7.39E-25
Lamivudine	0.25	1.74E-18	2.64E-20
Vilazodone	0.17	1.39E-16	1.90E-18
Hydrocortisone	0.02	9.09E-12	8.34E-14
Albuterol	1.31	1.66E-09	1.23E-11
Sodium Polystrene sulfonate	0.22	3.16E-06	1.54E-08
Ibuprofen	1.30	2.06E-05	8.81E-08
Ondasetron	1.30	7.36E-05	2.85E-07
Suvorexant	0.25	0.0001	3.96E-07
Omperazole	1.31	0.0001	4.18E-07
Trazodone	1.30	0.0002	6.23E-07
Abraxane	0.39	0.0003	9.30E-07
Oxycodone	1.24	0.0004	1.30E-06
Hydroxyzine	1.26	0.0023	6.45E-06
Modafinil	0.44	0.0027	7.40E-06
Dronabinol	0.42	0.0049	1.30E-05
Tenofovir	0.45	0.0144	3.51E-05

^*Relative contribution value > 1: Risk and Relative contribution value < 1: Protective; FDR_Q: false discovery rate adjusted Q value; Wilcoxon_p: P values with Wilcoxon.

Table 4.

Significant diagnoses identified through perturbation-based contribution analysis for predicting adverse events.

Name	RC	Wilcoxon_p	FDR_Q
Tobacco use disorder	1.40	6.03E-61	5.48E-62
Esophageal reflux	1.41	5.97E-52	3.73E-53
Asthma, unspecified type, unspecified	1.53	1.43E-45	6.51E-47
Abdominal pain, unspecified site	1.61	1.44E-37	5.55E-39
Personal history of tobacco use	1.37	2.96E-25	7.40E-27
Unspecified sleep apnea	1.53	2.79E-22	5.58E-24
Obesity, unspecified	1.57	9.85E-20	1.73E-21
Long-term (current) use of steroids	1.38	4.71E-19	7.60E-21
Other chronic pain	1.38	8.54E-15	1.03E-16
Personal history of colonic polyps	1.32	2.13E-14	2.48E-16
Lumbago	1.36	6.27E-14	6.68E-16
Myalgia and myositis, unspecified	1.35	2.19E-09	1.61E-11
Other and unspecified hyperlipidemia	1.30	4.39E-07	2.37E-09
Other specified counseling	0.32	4.54E-05	1.82E-07
Pain in joint, forearm	1.26	0.001	4.23E-06

^*Relative contribution value > 1: Risk and Relative contribution value < 1: Protective; FDR_Q: false discovery rate adjusted Q value; Wilcoxon_p: P values with Wilcoxon.

Table 5.

Significant SDoH identified for predicting adverse events risk.

Name	mean	sd	p	Impact on adverse events	Type of SDoH
Psychotherapy	−0.03	0.02	0.0005	Individuals undergoing trauma focused psychotherapy have lesser risk of adverse events	Individual
Park Proximity	−0.04	0.03	0.002	Individuals belonging to zip codes with households with high park proximity have lesser risk of adverse events	Neighborhood
CBT	−0.04	0.03	0.002	Individuals undergoing cognitive behavioral therapy have lesser risk of adverse events	Individual
Households with limited English-speaking household	0.04	0.03	0.002	Individuals belonging to zip codes with households with limited English speaking have higher risk of adverse events	Neighborhood
Income segregation	−0.05	0.04	0.003	Households with higher income segregation have higher risk of adverse events	Neighborhood
Normalized difference vegetation index	−0.03	0.03	0.004	Low vegetation/greenery have higher risk of adverse events	Neighborhood
Gini index	−0.04	0.03	0.004	Zip codes with low Gini index have higher risk of adverse events	Neighborhood
Veteran status	0.04	0.03	0.007	Veterans are at a higher risk of adverse events	Individual
Age	0.02	0.03	0.021	Younger patients have higher risk of adverse events	Neighborhood
Health literacy status	−0.02	0.03	0.039	Individuals belonging to zip codes with households with low health literacy status have higher risk of adverse events	Neighborhood

CI: confidence interval, SD: standard deviation.

5. Discussion

To enhance our deep learning model, DeepBiomarker2, for assessing adverse event risk in PTSD+AUD patients, we incorporated a wide range of features including lab test results, diagnosis, medication use, SDoH parameters, psychotherapy and veteran data. By incorporating input sequences and adopting a multimodal approach, we were able to effectively capture the temporal dynamics that exist between features from various domains. This led to substantial improvements in model performance (AUC score exceeding 0.93). To gain further insights into the specific biomarkers relevant to the intersection of PTSD+AUD and the associated adverse events, we have categorized our top biomarkers based on their types:

5.1. Lab-test results closely related to adverse events risk in PTSD+AUD

5.1.1. Inflammatory-based biomarkers

In our study, we have identified two inflammatory-based biomarkers: white blood cells and lymphocytes that could be valuable for assessing the risks of adverse events in patients with co-occurring PTSD and AUD. Exposure to traumatic stressors and psychological trauma is associated with adverse health outcomes and increased healthcare utilization. These exposures have been linked to various conditions, including cardiovascular disease, diabetes, gastrointestinal disease, fibromyalgia, and neuropsychiatric disorders. The biological pathways involving stress axes contribute to these disorders (JA., 2004). Incorporating inflammatory biomarkers in risk prediction models is essential for advancing research on mental disorders (Gill, 2008; Katrinli, 2021). There is evidence suggesting a potential connection between PTSD and autoimmune diseases, as well as dysregulation of the immune system in individuals with chronic PTSD and heavy drinking (Alexandrova, 2022). These inflammatory factors can impact the central nervous system and may serve as targets for future interventions and biomarkers for assessing symptom severity and suicide risk (Keaton, 2019).

5.1.2. Heme-based biomarkers

PTSD patients may have abnormal levels of biomarkers related to hematopoiesis, inflammation, endothelial function, and coagulability (Lindqvist D and Dhabhar 2017). Studies suggest that red blood cells (RBCs) can promote inflammation and atherosclerosis when exposed to oxidative stress. Alcohol use disorder can directly and indirectly affect hematological parameters, while chronic opioid use alters blood homeostasis (HS., 1997). Our findings show that opioid users have low RBC levels and increased MCV, HGB, HCT, and RDW compared to healthy individuals, Alcoholics also have low levels of hemoglobin, RBCs, hematocrit, and platelets, as well as elevated MCV, RDW, and MCH levels (Jain R and Narnoli 2020). Although the impact of these hematological biomarkers on PTSD and adverse events is not yet established, they do impact the quality of life in patients with comorbidities (Fan, 2019; Nagao, 2017; Wolf, 1989). Our study suggests that HGB, HCT, RDW, RBCs, MCHC, MCH and MCV biomarkers as potential indicators for adverse event risk in patients with both PTSD and AUD.

5.1.3. Kidney-based biomarkers

A study in patients with AUD found lower creatine levels, suggesting a decline in energy metabolism (Demnitz, 2020). Proteinuria, a marker of kidney damage, may not provide a definite diagnosis and should be evaluated alongside other kidney based biomarkers. Biomarkers such as chlorine, sodium, calcium, urea nitrogen, potassium, and carbon dioxide directly relate to acute renal insufficiency and chronic kidney disease (Gounden, 2023; HJ., 2021; Mehmood, 2021; Rein, 2019; Yamada, 2021). These conditions manifest as sudden onset of reduced urine output, acidosis, fluid imbalance, and electrolyte disorders (Goyal, 2023). Although the association between these biomarkers and PTSD+AUD is limited, our research has identified their potential for assessing the risks of adverse events in patients with both disorders.

5.1.4. Liver-based biomarkers

Our research has identified two biomarkers with potential for assessing the risks of adverse events in patients with co-occurring PTSD and AUD: ALT and albumin. Liver enzymes, such as ALT and AST, are often elevated in cases of liver injury, even in asymptomatic individuals (Malakouti, 2017). Psychological stress, including PTSD, has been associated with liver diseases and poor cardiovascular health (A., 2015; Osna, 2017). Biomarkers of systemic inflammation, ALT, AST and symptoms of PTSD have shown associations with the initiation and progression of atherosclerosis. Studies have also linked elevated AST levels to AUD, external-cause mortality (including suicide and accidents), higher cardiovascular risk factors, and non-liver related mortality (Conigrave, 2002). Serum albumin, a protein with antioxidant properties, has been examined in psychiatric diseases, with low levels observed in depressed patients and drug addicts (LIhua M et al. 2020; Michopoulos et al., 2016). However, further research is needed to fully understand the associations between these biomarkers and the risk of adverse events in patients with PTSD+AUD.

5.1.5. Metabolic disorder-based biomarker

PTSD may increase the risk of insulin resistance and diabetes through various mechanisms. Recent research has proposed: elevated inflammatory markers, alterations in the hypothalamic-pituitary-adrenal axis, and associated factors such as elevated body mass index, poor sleep and unhealthy lifestyle choices (Hirotsu, 2015). These factors contribute to the development of diabetes. Our results are in line with another study, that found biomarkers particularly in glucose metabolism, show promise for distinguishing controls from PTSD individuals and PTSD individuals with TBI (Xu, 2023). Also, opioid use, alcohol use and suicide all lead to dysregulation of glucose metabolism (Forsman, 2017; Ojo, 2018; Tomasi, 2019). Dysregulation of glucose metabolism appears to be an early event which precedes other clinical abnormalities, including systemic inflammation, dyslipidemia, and other cardiovascular diseases (Galicia-Garcia, 2020).

5.2. Effect of medication use on PTSD+AUD for adverse events prediction

5.2.1. Indicators of high risk

5.2.1.1. Hydrocodone and Oxycodone.

Hydrocodone and oxycodone are opioids used for pain relief, but they can cause side effects like constipation, respiratory depression, and dependence (Pathan, 2012). OUDs combined with AUD and PTSD can worsen mental and physical health (Borsheski, 2014). Treatment plans for patients with these dual diagnoses are crucial for better outcomes. A study found that individuals with OUD had increased cravings for alcohol and other substances, more severe PTSD symptoms, and higher levels of depression and anxiety (Peck, 2018). These patients may have difficulty following standard treatment, leading to frequent emergency department visits and negatively impacting their overall quality of life.

5.2.1.2. Omeprazole.

Omeprazole is a potent acid-suppressive medication used to treat various gastrointestinal disorder (Ahmed, 2019). However, recent studies have raised concerns about its potential association with increased risk of depression and anxiety. It is believed that omeprazole may dysregulate the human microbiome, affecting the microbiome-gut-brain axis and contributing to the development of anxiety and depression (Bushnell, 2022; Clapp, 2017). The interaction between omeprazole and the microbiome may lead to elevated levels of gastrin, further exacerbating the pathogenesis of depression and anxiety (Bruno, 2019). This can result in altered cytokine expression in the brain, impacting neurotransmission of serotonin and dopamine. Clinical studies have demonstrated higher depressive scores in patients treated with omeprazole, particularly with higher dosages (Laudisio, 2018). Alcohol consumption in individuals with gastroesophageal reflux disease (GERD) can increase reflux, and whole omeprazole can reduce reflux after alcohol consumption in healthy individuals, its effect on GERD patients required further investigation (Chen, 2010). Additional research is needed to understand the potential mechanisms linking omeprazole to depression and alcohol use.

5.2.1.3. Ibuprofen.

Ibuprofen is a commonly used non-steroidal anti-inflammatory drug (NSAID) with potential adverse effects such as liver toxicity, kidney toxicity, and stomach bleeding (Ghlichloo, 2023). Alcohol consumption can worsen the liver toxicity of NSAIDs, raising concerns about the interaction between ibuprofen and alcohol (Kim, 2021). Studies have shown that ibuprofen and alcohol can synergistically induce hepatoxicity, indicating caution in the use of ibuprofen in patients with alcohol use (J., 2009). However, preclinical studies suggest that ibuprofen may have potential benefits in the treatment of anxiety and PTSD.

5.2.1.4. Albuterol.

Studies propose that early administration of certain drugs can potentially prevent the development of PTSD, by interfering with the consolidation or retrieval of traumatic memories. Albuterol, a B2-adrenergic receptor agonist commonly used to treat asthma attacks and respiratory issues, has shown interesting effects (Kobayashi, 2011). It can rapidly cross the blood-brain barrier, and low-dose inhalation of nebulized albuterol enhances avoidance learning in rats, while high doses interfere with it (Elias, 2004). This suggests that albuterol may impact initial fear responses (McAllister, 1991). However, more research is needed, as the current studies have limitation, such as small patient samples.

5.2.1.5. Gabapentin.

Gabapentin is commonly prescribed for post-herpetic neuralgia and as adjunctive therapy for focal seizures (Yasaei, 2023). However, it is increasingly being used off-label for psychiatric conditions, as found in a recent analysis of US prescription practices (Fukada, 2012). While gabapentin has evidence supporting its use in conditions such as AUD, alcohol withdrawal, social anxiety disorder, and severe panic disorder, caution is advised regarding its use for bipolar disorder, major depressive disorder (MDD), PTSD, obsessive compulsive disorder (OCD), stimulant use disorder, or opioid withdrawal (Martin, 2022). Gabapentin may alleviate ongoing symptoms in patients with alcohol abuse, but there is a possibility that it may worsen SUD risks in patients with co-occurring PTSD and AUD.

5.2.1.6. Ondansetron.

Trauma exposure often leads to persistent somatic symptoms, such as nausea, dizziness, and headaches. Studies have linked increased violence exposure to more frequent somatic symptoms in adolescents and female military veterans. Interestingly, gastrointestinal (GI) symptoms, particularly in children exposed to violence, are prevalent (Escalona, 2004; Stein, 2004). A previous study found a significant association between postoperative nausea, vomiting and PTSD (Drews, 2015). However, the link between peri-traumatic nausea and subsequent PTSD symptom development remains unexplored. A comprehensive study used ondasetron administration as a surrogate for nausea, commonly prescribed in emergency departments for nausea. Ondansetron is a 5-hydroxytryptamine receptor (5HT3A R) antagonist. Subsequent analyses found that peri-traumatic nausea is a prospective predictor of PTSD symptom development and that ondansetron increases PTSD severity in these patients (Michopoulos, 2019). In a double-blind, placebo-controlled clinical trial, the efficacy of low-dose oral ondansetron in treating alcohol use disorder was assessed among individuals of European and African ancestry (Johnson, 2011). While a previous study suggested that ondansetron might be effective in reducing alcohol consumption, the results of another trial did not confirm those findings. The trial found that ondansetron did not significantly differ from the placebo in terms of alcohol consumption, including drinks per drinking day, heavy drinking days per week and drinks per day (Seneviratne, 2022). These outcomes were consistent across various genotypes, and no severe adverse events were reported during the study. However, the trial may have been underpowered to detect potential interactions between medication and genotype. The administration of ondansetron has shown potent inhibition of experimentally induced opioid dependence and withdrawal responses in mice and humans (Wu, 2023). However, its efficacy appears reduced or absent in individuals with chronic opioid use disorders, further studies are necessary to investigate ondansetron pharmacokinetics in PTSD+AUD patients.

5.2.1.7. Trazodone.

Trazodone, a medication used to manage major depressive disorder, raises concerns about its association with suicide and self-harm, particularly in individuals under 25 (Shin, 2023). It is an antidepressant that works by inhibiting both serotonin transporter and serotonin type 2 receptors. A meta-analysis of antidepressant trials found increased suicidal behavior risk in younger adults but not in those aged 25–64, and a reduced risk in those 65 or older (Dufort, 2021). However, the impact of different antidepressants, dose, and duration of use remains unclear. A study assessed the influence of various antidepressants on suicide and self-harm risk on individuals aged 20–64 (Coupland, 2015). When compared to zolpidem, the hazard of suicide attempts was 61% higher with trazodone, suggesting caution in its use (Lavigne, 2019).

5.2.1.8. Hydroxyzine.

Generalized anxiety disorder (GAD), is a prevalent, chronic psychiatric condition often managed in primary care, with treatment options encompassing cognitive behavioral therapy and pharmacological interventions (Guaiana, 2010). Among these medications are antidepressants, benzodiazepines and hydroxyzine. However, hydroxyzine is associated with an elevated risk of suicidal thoughts and behaviors but its mechanism remains unclear (Robertson, 2009; Tubbs, 2021). Our study sought to draw attention to explore the relationship between hydroxyzine and suicidal ideations and actions.

5.2.2. Indicators of low risk

5.2.2.1. Vilazodone.

Vilazodone is a dual-action antidepressant that inhibits serotonin transporters and partially activates serotonin-1a (5-HT1A) receptors (Singh, 2012). This combined mechanism enhances serotonin facilitation in the brain, acting as a serotonin partial agonist and reuptake inhibitor. The partial agonist action on 5-HT1A receptors may potentially reduce sexual dysfunction (GW., 2005; R., 1998). Another medication, buspirone, which also acts as a 5-HT1A receptor partial agonist, has shown effectiveness in reducing anxiety and cravings in patients with alcohol dependence (Trevedi, 2006). Utilizing vilazodone for comorbid depression and alcohol dependence may be advantageous due to shared pathways involved in these conditions.

5.2.2.2. Dronabinol.

Dronabinol is a synthetic form of THC, the psychoactive component of marijuana, and it has demonstrated antidepressant effects in patients with depressive disorders (Ng T, 2023). It interacts with cannabinoid 1 (CB1) receptor and regulates levels of 2-arachidonyl glycerol (2-AG) and anandamide, depending on the type of depression (Hill, 2008). Dronabinol has shown low abuse potential, reduced withdrawal symptoms and cravings, and decreases the reinforcing effects of cannabis, promoting abstinence (Levin, 2011). However, these effects have mainly been observed in laboratory settings among non-treatment seeking cannabis users. Clinical trials have reported positive subjective effects and improved treatment retention with dronabinol (Connor, 2021), but further research is needed, particularly in high-risk individuals with comorbid PTSD and AUD, higher doses, combination therapies, or unconventional interventions.

5.2.2.3. Suvorexant.

Suvorexant, a dual orexin receptor antagonist, improves sleep by blocking the orexin receptors OX1R and OX2R (Hellmann, 2020). The orexin system is involved in various physiological functions, including the behavioral response to alcohol (Koob, 2016; Mahler, 2012). Studies have shown that orexin receptors play a role in alcohol motivation and craving, and dual orexin receptor antagonists can reduce alcohol consumption in animal models (Anderson, 2014). Suvorexant can target both alcohol-related symptoms and comorbid conditions such as anxiety, insomnia, and addiction. Clinical studies have demonstrated the involvement of orexin system in emotional dysregulation during alcohol withdrawal (Mogavero, 2023). Suvorexant has been found to promote sleep, reduce alcohol consumption, and increase abstinence in alcohol-exposed groups (Campbell, 2020). Future clinical trials should explore the effects of suvorexant on sleep-related outcomes, alcohol use, and other neuropsychiatric symptoms to address the comprehensive symptomatology of alcohol use disorder.

5.2.2.4. Lamivudine.

Lamivudine, used for Hepatitis B and Acquired Immune Deficiency Syndrome (AIDS)(2019 Apr 18]. Available from: https://www.ncbi.nlm.nih.gov/books/NBK548553/). It protects the central nervous system and repairs neuronal damage cause by alcohol. It regulates alcohol concentration and activates acetaldehyde dehydrogenase (ALDH) to metabolize acetaldehyde. Lamivudine improves AUD symptoms, including alcohol tolerance and sobering time, with limited neuropsychiatric effects (Han, 2022).

5.2.2.5. Tenofovir.

Tenofovir, used for Hepatitis B and AIDS, has shown benefits in reducing PTSD symptoms in HIV patients (Wassner, 2020). AUD is common in HIV patients and can hinder treatment adherence. Frequent alcohol use is associated with increased viral loads and reduced CD4+ T cell levels (Edelman, 2012; Molina, 2014). While the direct benefit of tenofovir on adverse events risk in PTSD+AUD is not well-established, reducing alcohol consumption is crucial for managing co-morbidities.

5.2.2.6. Modafinil.

Modafinil is a stimulant that is traditionally used for sleep disorders and cognitive enhancement. It has gained recognition for its anti-inflammatory, anti-oxidative and neuroprotective properties. One study suggests that modafinil has the potential to mitigate neuro-inflammation associated with traumatic brain injury, potentially reducing TBI-related mortality and morbidity (Ozturk, 2023). Another study, found modafinil had rapid improvement in a patient’s post operative consciousness following brain surgery and also highlighted its effectiveness in enhancing wakefulness and responsiveness in patients with central nervous system trauma (VM., 2005). Additionally, an examination of dopaminergic medications, including bromocriptine, dihydrexidine, modafinil, rotigotine and pramipexole, in a PTSD mouse model reveals their antidepressant-like activity, supporting the potential of dopaminergic D2/D3 receptor activation as a promising avenue for PTSD pharmacology (Malikowska-Racia, 2019). Poor impulse control is a critical factor in the development and progression of substance use disorders. A randomized, double-blind, placebo-controlled study done in alcohol dependent patients found improved self-reported state impulsivity but no effect on abstinent and heavy drinking days (Joos, 2013). Another study found modafinil to promote abstinence in cocaine dependent patients over a span of 12 weeks with a 4 week follow up period. Patients were randomized into three groups: placebo, modafinil 200 mg and modafinil 400 mg. While no significant differences emerged in average weekly cocaine non-use days between modafinil and placebo for the entire sample, modafinil 200 mg did show promise in increasing the maximum consecutive non-use days and reducing cravings (Anderson, 2009). Additionally, post hoc analysis revealed that modafinil was particularly effective in elevating the percentage of non-use days in participants without a history of alcohol dependence. Overall, these findings suggest that modafinil, when combined with individual behavioral therapy, can be effective in increasing non-use days among cocaine-dependent individuals without co-occurring alcohol dependence, as well as in reducing cocaine cravings. Additionally, within our patient cohort, those who underwent cognitive behavioral therapy demonstrated superior outcomes compared to those who did not receive this therapy.

Other drugs such as sodium polystyrene sulfonate (SPS) and abraxane were found to decrease adverse events risk in PTSD+AUD patients. SPS is a medication used to treat hyperkalemia (Rahman S et al.) and enhances lysosomal function, mitigating proteotoxicity driven by amyloid-B and hyper-phosphorylated-TAU in Alzheimer’s. (Arputhasamy, 2023). While abraxane, an albumin-bound nanoparticle formulation, used for glioblastoma (GBM) treatment (Gould, 2021; Zhang, 2020) had improved cognitive function, prevented injury-induced brain abnormalities, and increased brain glucose metabolism (Cross, 2019). While their impact on decrease in adverse events risk in PTSD+AUD is yet to be studied, further research is warranted for testing these drugs in preclinical and clinical models of PTSD+AUD to evaluate its real-world potential to reduce adverse events.

5.3. Effect of multiple SDoH parameters on adverse events prediction

5.3.1. SDoH parameters that are positively correlated

5.3.1.1. Veterans.

Alcohol use disorder is a significant public health issue in the US, particularly affecting veterans. We found that veterans have higher incidences of adverse events risks as opposed to civilians. Our results are in line with a study that showed veterans with lifetime AUD had higher rates of psychiatric disorders and suicidal behavior, with increased severity of AUD (Panza, 2022). Factors associated included younger age, male gender, white ethnicity, unmarried, retired, having experienced adverse childhood events, trauma, and lifetime PTSD. Our study emphasizes the high prevalence of PTSD+AUD in US veterans and its connection to significant adverse events risk. This also highlights the need for comprehensive screening, prevention, and interventions targeting both alcohol use and psychiatric concerns.

5.3.2. SDoH parameters that are negatively correlated

5.3.2.1. Cognitive behavioral therapy and Trauma focused psychotherapy.

These are psychotherapeutic approaches that help to identify and transform harmful behavior and emotional patterns, making it particularly effective in improving problem-solving and coping skills during stressful times (Nakao, 2021). Although real world data is limited, our study found that PTSD+AUD patients receiving these therapies have a lower risk of adverse events compared to those who do not undergo these treatments.

Other SDoH parameters such as individuals residing in zipcodes with households not located near parks, higher income segregation, limited vegetation and greenery, low Gini index, patients belonging to zip codes with households that have limited English-speaking capacity, younger patients and low health literacy status were found to exhibit elevated incidences of PTSD, suicide related events, opioid use and AUD.

It is crucial to consider such multidimensional aspects when translating these results into routine clinical practice to effectively address the challenge of existing health disparities.

5.4. New hypotheses generated by DeepBiomarker2 on adverse events risk in PTSD+AUD

Previous studies on PTSD+AUD often overlooked multi-dimensional biomarkers as contributors to adverse events. In this study, we propose two hypotheses exploring interdisciplinary indicators of mental health disorders and adverse events. One study found a link between high immunoglobulin-E and WBC levels and worsening depressive scores in bipolar patients during high pollen seasons (Xiong W et al. 2017). We propose a hypothesis linking allergies to adverse events in PTSD+AUD, suggesting inflammatory mediators activated by asthma and allergic rhinitis as potential biomarkers. Additionally, we explore the association between elevated AST levels and increased external cause mortality, suicide, and injury. Elevated ALT and AST levels are closely tied to liver fat accumulation and associated health risks (Sohn, 2015). We hypothesize that these disorders coincide with brain dysfunction, depression, and cognitive decline due to inflammation, hypoperfusion, and neuronal damage. These findings open avenues for further research and personalized treatment approaches for this debilitating illness.

5.5. Utilizing data-driven models to guide the mitigation of adverse events in PTSD+AUD patients

Our study utilized EMR (both structured and unstructured) and non-EMR data to propose a novel assessment tool on risk of adverse events among PTSD+AUD patients. PTSD+AUD is a complex disorder influenced by genetics, neurobiology, environment, and psychology. Integrating real world data is crucial for personalized treatment strategies and predicting adverse events risks. Our goal was to develop a practical algorithm for routine care, considering comprehensive multimodal factors. Our study’s strengths include a large sample size, convenience, affordability, relevant tests, multi-dimensional information, and practical applicability. Although consensus on optimal biomarker combinations is lacking, our study provides a foundation for identifying reliable biomarkers and building consensus in clinical use.

5.6. Limitation of our study

Our study has limitations that should be considered. Inconsistencies in biochemical test results and limited representation of certain lab tests in our database may introduce enrollment bias and affect statistical power. The use of EMR data from a specific timeframe may be influenced by changes in treatment practices. “Furthermore, our study utilized ACS data to derive neighborhood-level SDoH features aggregated at the zip code level. The selection of the zip code as the most appropriate geographical unit was necessitated by the utilization of the EMR data, which exclusively provides 5-digit zip codes. Additionally, due to the necessity of anonymizing the data before gaining access, we didn’t have the ability to locate the individual to a more specific region. “While our EMR data spanned a 16-year period, the specific timeframe of the ACS data used were 5-year estimates, potentially introducing a temporal gap in the representation of neighborhood characteristics.” “Although zip codes provide a convenient and commonly used means of data aggregation, they are not without limitations. Variations in physical size and the potential masking of trends over time due to the 5-year estimates period are among these limitations. Alternatives, such as census tracts or other data sources, might offer more nuanced insights.” Causal interpretations require caution, and future randomized clinical trials or prospective designs are needed. Comorbidities had a greater influence than lab tests, suggesting the need for further exploration. Inconsistencies and missing information in SDoH data may have affected accuracy. The reliance on neighborhood level SDoH parameters without individual level variables like income limits the assessment of socioeconomic factors. Future studies should include individual-level SDoH information for a more comprehensive understanding. Our future research will involve multiple datasets, refined algorithms, informative biomarkers and application of advanced deep learning models to address these limitations and enhance the applicability of our findings.

6. Conclusion

Our personalized approach has the potential to enhance prevention efforts and mitigate the impact of adverse events in PTSD+AUD patients. Based on our results, we identified several medications including vilazodone, dronabinol, tenofovir, suvorexant, modafinil, and lamivudine, which have the potential to reduce the risk of adverse events among PTSD+AUD patients. While universal prevention programs may yield benefits in the current landscape, the insights derived from DeepBiomarker2 offer valuable and refined information that can be utilized to design and develop personalized prevention and intervention programs, which are programs that can be tailored to address health disparities prevalent among these high-risk patients. Moving forward, further research should focus on utilizing this knowledge to effectively address the unique challenges faced by these individuals.

Appendix A. Diagnosis codes

1. PTSD:

309.81, F43.10, F43.11, F43.12.

2. ASUD:

291.0, 291.1, 291.2, 291.3, 291.4, 291.5, 291.8, 291.81, 291.82, 291.89, 291.9, 292.0, 292.11, 292.12, 292.2, 292.81, 292.82, 292.83, 292.84, 292.85, 292.89, 292.9, 357.5, 425.5, 535.30, 535.31, 571.0, 571.1, 571.2, 571.3, 648.30, 648.31, 648.32, 648.33, 648.34, 965.00, 965.01, 965.02, 965.09, 968.5, 969.6, E850.0, E854.1, E860.0, E935.0, E938.5, E939.6, V654.2, 303, 303.0, 303.00, 303.01, 303.02, 303.03, 303.9, 303.90, 303.91, 303.92, 303.93, 304, 304.0, 304.00, 304.01, 304.02, 304.03, 304.1, 304.10, 304.11, 304.12, 304.13, 304.2, 304.20, 304.21, 304.22, 304.23, 304.3, 304.30, 304.31, 304.32, 304.33, 304.4, 304.40, 304.41, 304.43, 304.5, 304.50, 304.51, 304.52, 304.6, 304.60, 304.61, 304.62, 304.63, 304.7, 304.70, 304.71, 304.72, 304.73, 304.8, 304.80, 304.81, 304.82, 304.83, 304.9, 304.90, 304.91, 304.92, 304.93, 305, 305.0, 305.00, 305.01, 305.02, 305.03, 305.1, 305.10, 305.12, 305.13, 305.2, 305.20, 305.21, 305.22, 305.23, 305.3, 305.30, 305.31, 305.33, 305.4, 305.40, 305.41, 305.42, 305.43, 305.5, 305.50, 305.51, 305.52, 305.53, 305.6, 305.60, 305.61, 305.62, 305.63, 305.7, 305.70, 305.71, 305.72, 305.73, 305.8, 305.80, 305.81, 305.83, 305.9, 305.90, 305.91, 305.92, 305.93, F10.10, F10.11, F10.120, F10.121, F10.129, F10.14, F10.151, F10.159, F10.180, F10.188, F10.19, F10.20, F10.21, F10.220, F10.221, F10.229, F10.230, F10.231, F10.232, F10.239, F10.24, F10.250, F10.251, F10.259, F10.26, F10.27, F10.280, F10.288, F10.29, F10.920, F10.921, F10.929, F10.94, F10.951, F10.959, F10.96, F10.97, F10.980, F10.982, F10.988, F10.99, F11.10, F11.11, F11.120, F11.121, F11.129, F11.14, F11.159, F11.188, F11.19, F11.20, F11.21, F11.220, F11.221, F11.222, F11.229, F11.23, F11.24, F11.250, F11.259, F11.288, F11.29, F11.90, F11.921, F11.929, F11.93, F11.94, F11.988, F11.99, F12.10, F12.11, F12.121, F12.122, F12.129, F12.150, F12.151, F12.159, F12.180, F12.188, F12.19, F12.20, F12.21, F12.220, F12.23, F12.250, F12.259, F12.288, F12.29, F12.90, F12.920, F12.921, F12.922, F12.929, F12.959, F12.980, F12.988, F12.99, F13.10, F13.11, F13.129, F13.14, F13.180, F13.188, F13.19, F13.20, F13.21, F13.220, F13.221, F13.229, F13.230, F13.231, F13.232, F13.239, F13.24, F13.259, F13.27, F13.280, F13.29, F13.90, F13.920, F13.921, F13.929, F13.930, F13.931, F13.939, F13.94, F13.97, F13.980, F13.99, F14.10, F14.11, F14.120, F14.121, F14.122, F14.129, F14.14, F14.151, F14.159, F14.180, F14.182, F14.188, F14.19, F14.20, F14.21, F14.220, F14.221, F14.222, F14.229, F14.23, F14.24, F14.250, F14.251, F14.259, F14.280, F14.282, F14.288, F14.29, F14.90, F14.920, F14.921, F14.929, F14.94, F14.951, F14.959, F14.980, F14.988, F14.99, F15.10, F15.11, F15.121, F15.129, F15.14, F15.159, F15.180, F15.188, F15.20, F15.21, F15.220, F15.222, F15.229, F15.23, F15.259, F15.29, F15.90, F15.920, F15.921, F15.929, F15.93, F15.94, F15.950, F15.951, F15.959, F15.980, F15.982, F15.988, F15.99, F16.10, F16.11, F16.129, F16.159, F16.20, F16.21, F16.221, F16.24, F16.259, F16.283, F16.90, F16.921, F16.929, F16.950, F16.959, F16.980, F16.983, F16.988, F16.99, F17.200, F17.201, F17.203, F17.208, F17.209, F17.210, F17.211, F17.213, F17.218, F17.219, F17.220, F17.223, F17.228, F17.229, F17.290, F17.298, F17.299, F18.10, F18.11, F18.19, F18.20, F18.24, F18.90, F19.10, F19.11, F19.120, F19.121, F19.129, F19.14, F19.150, F19.159, F19.180, F19.181, F19.188, F19.19, F19.20, F19.21, F19.221, F19.229, F19.230, F19.231, F19.232, F19.239, F19.24, F19.259, F19.280, F19.29, F19.90, F19.920, F19.921, F19.922, F19.929, F19.930, F19.931, F19.939, F19.94, F19.950, F19.951, F19.959, F19.96, F19.980, F19.982, F19.988, F19.99.

3. SRE:

V62.84, R45.851, E950.3, E956, E950.4, E950.0, E958.8, T14.91, E950.9, E958.9, T14.91XA, E950.5, E950.2, E953.0, E958.1, E953.8, E950.1, E950.7, E952.0, E958.0, E957.1, E957.0, E958.5, E952.1, E955.4, T14.91XD, E950.6, E953.9, E955.0, E957.9, E958.7, E958.3, E954, T14.91XS, E951.0, E951.8, E952.8, E953.1, E955.1, E958.6, E958.2, E955.9, E955.2, X83.8XXA, T42.42 A, T43.592 A, T39.1×2A, X78.8XXA, X78.9XXD, T42.6×2A, X78.9XXA, T43.222 A, T50.902 A, X83.8XXD, T39.312 A, T43.212 A, T45.0×2A, X78.1XXA, X78.8XXD, T50.992 A, T40.2×2A, T43.292 A, T43.012 A, T39.012 A, T42.8×2A, X78.0XXA, T51.0×2A, T40.5×2A, T40.4×2A, T40.1×2A, T44.7×2A, T38.3×2A, T44.6×2A, T48.1×2A, T46.5×2A, T71.162 A, T48.3×2A, T43.022 A, T44.3×2A, T50.902D, X79. XXXA, T65.92XA, X78.1XXD, T51.92XA, T42.1×2A, T65.892 A, T56.892 A, T43.622 A, X80. XXXA, T42.4×2D, X78.0XXD, T42.72XA, T43.3×2A, X76. XXXD, T48.4×2A, T51.2×2A, T46.4×2A, T39.1×2D, T40.7×2A, T54.92XA, T40.602 A, T45.512 A, X76. XXXA, T43.222D, T39.392 A, T47.1×2A, T50.902 S, X74.9XXD, T39.092 A, T38.1×2A, X74.9XXA, T39.312D, T38.892 A, T43.612 A, X82.8XXA, T42.6×2D, T43.632 A, T46.1×2A, T45.0×2D, T50.992D, T54.2×2A, T40.3×2A, T39.012D, T43.4×2A, T58.02XA, T43.592D, X81.0XXA, T43.202 A, T43.8×2A, T44.992 A, T45.2×2A, T40.1×2D, T41.292 A, T50.2×2A, T48.6×2A, T50.7×2A, T49.0×2A, T46.3×2A, T42.0×2A, T36.1×2A, T36.0×2A, X74.9XXS, X72. XXXD, T43.012D, T51.8×2A, T51.0×2D, T54.92XS, T54.3×2A, T65.892D, T65.92XD, T65.222D, T50.3×2A, T48.5×2A, T47.0×2A, T46.6×2A, T65.92XS, T54.1×2A, T52.4×2A, T52.0×2A, T55.1×2A, T59.892 A, T42.4×2S, T42.3×2A, T36.3×2A, T37.8×2A, T38.2×2A, T38.3×2D, T40.992 A, T40.8×2A, T44.4×2A, T43.692 A, T45.2×2D, T44.7×2D, T43.502 A, T71.192 A, X79. XXXD, X83.2XXA, X83.8XXS, X72. XXXA, X71.9XXA, T48.202 A, T40.2×2D, X74.8XXS, T48.3×2D, T39.8×2A, T47.4×2A, T47.6×2A, T50.6×2A, T49.6×2D, T43.3×2D, T50.5×2A, X74.01XA, X73.0XXA, T49.6×2A, X72. XXXS, X78.9XXS, T39.92XA, X80. XXXD, X81.8XXA, T39.4×2A, X77.8XXA, T50.2×2D, T43.622D, T43.292D, T45.4×2A, T46.0×2A, T41.3×2A, T42.5×2A, T42.6×2, T46.7×2A, T46.8×2A, T46.5×2D, T43.1×2A, T43.92XA, T40.5×2D, X71.0XXS, X71.3XXA, X71.8XXA, T43.212D, T46.2×2A, T40.8×2D, T40.602D, T43.022D, T44.1×2A, T46.4×2D, T65.222 S, T62.0×2A, T71.162D, T51.1×2A, T51.2×2D, T51.2×2S, T52.8×2A, T51.92XD, T50.8×2A, T56.892D, T58.92XA, T54.3×2S, T54.3×2D, T54.0×2A, T55.0×2A, T36.0×2D, T36.4×2A, T38.5×2A, T36.8×2A, T37.5×2A.