Skip to main content

Predicting the presence of depressive symptoms in the HIV-HCV co-infected population in Canada using supervised machine learning

Abstract

Background

Depression is common in the human immunodeficiency virus (HIV)-hepatitis C virus (HCV) co-infected population. Demographic, behavioural, and clinical data collected in research settings may be of help in identifying those at risk for clinical depression. We aimed to predict the presence of depressive symptoms indicative of a risk of depression and identify important classification predictors using supervised machine learning.

Methods

We used data from the Canadian Co-infection Cohort, a multicentre prospective cohort, and its associated sub-study on Food Security (FS). The Center for Epidemiologic Studies Depression Scale-10 (CES-D-10) was administered in the FS sub-study; participants were classified as being at risk for clinical depression if scores ≥ 10. We developed two random forest algorithms using the training data (80%) and tenfold cross validation to predict the CES-D-10 classes—1. Full algorithm with all candidate predictors (137 predictors) and 2. Reduced algorithm using a subset of predictors based on expert opinion (46 predictors). We evaluated the algorithm performances in the testing data using area under the receiver operating characteristic curves (AUC) and generated predictor importance plots.

Results

We included 1,934 FS sub-study visits from 717 participants who were predominantly male (73%), white (76%), unemployed (73%), and high school educated (52%). At the first visit, median age was 49 years (IQR:43–54) and 53% reported presence of depressive symptoms with CES-D-10 scores ≥ 10. The full algorithm had an AUC of 0.82 (95% CI:0.78–0.86) and the reduced algorithm of 0.76 (95% CI:0.71–0.81). Employment, HIV clinical stage, revenue source, body mass index, and education were the five most important predictors.

Conclusion

We developed a prediction algorithm that could be instrumental in identifying individuals at risk for depression in the HIV-HCV co-infected population in research settings. Development of such machine learning algorithms using research data with rich predictor information can be useful for retrospective analyses of unanswered questions regarding impact of depressive symptoms on clinical and patient-centred outcomes among vulnerable populations.

Peer Review reports

Background

With shared modes of transmission, co-infection of human immunodeficiency virus (HIV) and hepatitis C virus (HCV) is common, with approximately 2.3 million co-infected individuals worldwide [1, 2]. Depression is the most common neuropsychiatric manifestation among people living with HIV and those with chronic HCV. The prevalence of diagnosed clinical depression is two to fourfold higher among people living with HIV than the general population and reported to be as high as 24% among those with chronic HCV infection [3, 4]. Potential biological mechanisms include direct infection of the central nervous system and peripheral immune responses which have been shown to induce depression [3, 5]. Psychosocial risk factors including stigma, discrimination, lack of support, and substance use have also been shown to be contributory [3, 5]. Studies report an even higher depression prevalence in the co-infected population, which may be due to the co-existence of risk factors [6].

The presence of significant depressive symptoms may have an impact on outcomes in patients, even in the absence of a clinical depression diagnosis; for example, the presence of depressive symptoms is associated with non-adherence to antiretroviral therapy among people living with HIV, which may lead to increased viral load and suppressed immune function [7, 8]. Screening tools can be used to assess presence and severity of depressive symptoms and identify those at risk for major depression [9, 10], permitting early intervention. Co-infected individuals often live with multiple co-morbid conditions and thus spend a considerable amount of time in healthcare settings [11]. Depression screening among HIV and HCV infected patients is seldom routinely performed during clinical assessments or in longitudinal cohort studies, despite the known high prevalence of depression [12,13,14]. Multiple demographic, clinical and behavioural characteristics have been documented as risk factors for depression [15] and these data are generally collected in clinical and research cohorts. Thus, such data could be used retrospectively to predict the presence of depressive symptoms severe enough to be associated with negative health outcomes or an increased risk of being diagnosed with major depression and to follow this risk over time. Such measures will be useful for exploring important questions regarding depressive symptoms, their evolution and response to therapies in the co-infected population.

Machine learning includes robust techniques that enable accurate outcome predictions in medical research. In mental health research, machine learning has been used to predict current or future onset, disease course and treatment outcomes for psychiatric disorders including depression, anxiety, and schizophrenia [1, 16, 17]. A wide range of data sources have been used for developing these prediction algorithms including electronic medical records, neuroimaging, and social media. Demographic and clinical data have been used to create depression prediction models in the elderly and people with diabetes [18, 19]. However, similar models have not yet been developed in people living with HIV-HCV co-infection.

We leveraged a non-parametric supervised machine learning technique using cohort data to develop classification algorithms to predict the presence of depressive symptoms indicative of a risk for clinical depression and characteristics important for prediction of depressive symptoms in HIV-HCV co-infected individuals in Canada.

Methods

Data sources and study sample

We used data from the Canadian HIV-HCV Co-Infection Cohort (CCC), an open multicenter prospective cohort study, ongoing since 2003 and an associated sub-study, the Food Security and HIV-HCV co-infection study (FS sub-study) [20, 21]. The CCC recruits from 18 HIV centers, both urban and semi-urban across six Canadian provinces (Quebec, British Columbia, Alberta, Ontario, Nova Scotia, and Saskatchewan) [20]. Eligibility criteria include ≥ 16 years of age, documented HIV infection, and evidence of HCV infection (HCV RNA positive and/or HCV seropositive). The study had recruited 2018 participants as of July 2020. Participants are followed longitudinally, with follow-up visits every six months. Sociodemographic, behavioural, and health-related quality of life (HR-QoL) data are collected from participants by a standardized self-administered questionnaire at each visit. HR-QoL is measured using EuroQol-5 Dimension-3 Level (EQ-5D-3L) [22]. Clinical data including HIV/HCV treatment, co-morbidities, psychiatric diagnoses, and other medications are collected via medical chart reviews. Laboratory testing at each visit include HIV and HCV related tests, hematology, biochemistry, and liver profiles.

The FS sub-study is a mixed methods study conducted within the CCC between 2012 and 2015. All CCC participants were invited to participate and study visits were integrated into the biannual CCC visits. The FS sub-study recruited 725 participants and they were followed up for a maximum of 5 visits. The study collected data on food insecurity, general and mental health (including depression screening), treatment adherence and health care utilization using a self-administered questionnaire [21].

Depression screening was performed only in the FS sub-study; thus, the analytic sample in this study only included FS sub-study participant visits. The FS sub-study visits were merged with corresponding CCC visit data. As the two study visits for CCC and FS sub-study were, on occasion, not on the same day, information from visits within 3 months of each other were considered ‘concurrent’. We used three exclusion criteria to create the final study sample—1) participant visits were excluded if no depression screening measure (see below) was available at that visit; 2) participant visits were excluded if there was no corresponding CCC visit (within the 3-month window), as the predictors used in this analysis were derived from the CCC; and 3) all visits for a participant were excluded if no data was available for a predictor in all of their study visits.

Outcome

Depression screening was conducted in the FS sub-study using the Center for Epidemiologic Studies Depression Scale-10 (CES-D-10), which is a shortened version of the CES-D-20 scale [23]. The CES-D-10 is a 10-item Likert scale questionnaire that assesses presence and severity of depressive symptoms in the past one week. Each item is measured on a 4-point scale, with reverse scoring for the 2 positive items and a total score range of 0–30. We dichotomized the score at 10 to create the CES-D-10 classes (1/0), as a score ≥ 10 is widely considered for the presence of depressive symptoms indicative of high risk for clinical depression, hereafter referred to as depressive symptoms for brevity [23]. Both the scale and the dichotomization at 10 have been validated in HIV populations in Canada [24].

Predictors

We selected candidate predictors (x) from the CCC data based on the literature and subject matter expertise. We included predictors from five major categories: questions related to mental health, HR-QoL, sociodemographic, behavioural, and clinical characteristics; see Table 1. We selected a total of 137 candidate predictors, of which 136 were categorical and 1 was continuous (EQ-5D-3L—health state). From this list of candidate predictors, we selected a subset of predictors (x = 46) that may be more regularly available in most research settings based on expert opinion. See Table S1 in Appendix A for an exhaustive list of candidate predictors (x = 137) and their corresponding categories.

Table 1 Candidate predictors used in the random forest algorithms

Statistical analysis

Primary analysis

We assessed proportion of missing data for each predictor. For predictors with < 5% missing data, we carried the value from the last visit forward. For predictors with ≥ 5% missing data and when data were missing for a predictor for all visits for a participant, we used an additional category of “no response” for categorical variables, which we hypothesized could be informative in the prediction algorithm (See Table S1 in appendix A) since the participant decision to respond (or not) is itself potentially clinically informative.

We used the supervised machine learning technique of Random Forests (RF), an ensemble learning approach which uses bootstrap aggregation of multiple decision trees, combining predictions from these many trees [25]; see more details about RF in Appendix B. We used probability machines to estimate the CES-D-10 class probabilities at each visit and then determined CES-D-10 class at default probability threshold [26]. We developed two RF algorithms—1. Full algorithm: Using all candidate predictors (x = 137) and 2. Reduced algorithm: Using a selected subset of more commonly available predictors (x = 46) based on expert opinion, which could be more generalizable to research studies beyond the CCC; see appendix A. We split the analytical sample into training and testing data, using the recommended 80:20 split, such that we had data for performance evaluation (testing) that was completely independent of data used for model development (training) and thus ensure an unbiased evaluation [27]. We performed the 80/20 split using the “createdatapartition” function from the Caret package in R, such that both CES-D-10 classes were represented in each set [28]. The algorithm was then developed by tenfold cross-validation using only the training data and RF hyperparameters (i.e., various RF settings like number of decision trees) were tuned to maximize accuracy [27]; see Appendix C for details.

Additional analyses

We conducted several analyses to provide additional details about the classification characteristics from the main analysis and assess robustness of the results: A) Using one visit per individual (total 717 visits), to assess difference in performance compared to the use of multiple visits per individual; B) Algorithms using three different CES-D-10 thresholds—8, 13, and 15—based on suggested cut-offs in the literature [29, 30]; C) An algorithm that included food insecurity as a predictor, which was collected only in the FS sub-study. Food insecurity in the past 6 months was measured using a 10-item adult scale of the Household Food Security Survey Module (HFSSM) [31]. A categorical variable was used, with participants with 0–1, 2–5, or ≥ 6 affirmative responses classified respectively as being food secure, moderately food insecure, or severely food insecure, as per the Health Canada criteria; and D. RF regression algorithms to predict continuous CES-D-10 score, evaluated using R-squared and root mean squared error (RMSE), which provides information regarding differences between the predicted scores and the actual scores [32].

Performance evaluation

The final tuned algorithms were implemented in the testing data, which was not used in the development stage. The tuning parameters are shown in Table S2 in Appendix C. The overall performance and calibration measures are described in detail in Appendix C. To assess the ability to distinguish between classes (discrimination), we plotted receiver operating characteristic (ROC) and estimated the area under the ROC curve (AUC) [32, 33]. We used the default probability threshold of 0.50 for classification and at this threshold, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), positive likelihood ratio (LR +) and negative likelihood ratio (LR-) measures were then estimated with the 95% confidence intervals (CI) [34]. Finally, the RF importance metrics were generated, and importance plots were generated to present the 25 most important predictors in classifying participants with depressive symptoms by the two algorithms. We used RStudio v.1.2 and Stata v.16.0 to develop and evaluate these algorithms [35, 36]. The development of the RF algorithms was done using R package ranger and caret; for performance evaluation, we used the performance assessment function in R by Wong et. al. (2019) [28, 37,38,39].

Results

Study population

Of the 1973 FS sub-study visits in a total of 725 participants, 39 study visits were excluded based on the exclusion criteria described in the methods—16 visits (2 participants) with no CES-D-10 score, 18 visits (5 participants) with no con-current CCC visit and all 5 visits from 1 participant with no predictor data (EQ-5D health state) in all visits. Thus, 717 participants with a total of 1934 visits contributed to the final study sample. The participant characteristics at the first visit included in the sample are described in Table 2. The median CES-D-10 score was 10 (IQR, 5, 15), with 53% of the participants reporting the presence of depressive symptoms with CES-D-10 scores ≥ 10; 45% were prescribed one or more psychotropic medications such as bupropion and citalopram at baseline, but only 10% had a diagnosis of depression documented in their medical chart. Participants were predominantly male (73%) and white (76%). The population was vulnerable in terms of socioeconomic status (SES) characteristics with 73% unemployed, 76% with monthly income < $1500, 52% with high school being the highest level of education and 46% receiving welfare at baseline. Approximately 34% were current injection drug users, 62% current alcohol drinkers and 75% were current tobacco smokers. Only a small proportion had advanced liver disease (4%) or a current AIDS related illness (4%) and 35% were asymptomatic with a current CD4 cell count > 500 cells/μl (CDC clinical staging—A1).

Table 2 Baseline characteristics of participants in the study sample (n = 717)

Performance evaluation

The training data consisted of 1548 visits and testing data consisted of 386 visits. The algorithms in the primary analysis showed acceptable calibration, as seen in Figure S1 in Appendix C. With regard to discrimination, the ROC curve for the primary analyses is shown in Fig. 1, with the curve close to the upper left-hand corner. The estimated AUCs wert 0.82 (95% CI: 0.78–0.86) and 0.76 (95% CI: 0.71–0.81) for the full and reduced algorithms respectively. The estimated sensitivity, specificity, PPV, NPV, LR + and LR- are presented in Table 3. The importance plots with 25 most important predictors for both algorithms are shown in Fig. 2. Employment, HIV clinical stage, revenue source, body mass index (BMI), and education were the 5 most important predictors.

Fig. 1
figure 1

Receiver Operating Characteristic (ROC) curve for the A Full algorithm (x = 137) and B Reduced algorithm (x = 46)

Table 3 Performance evaluation in the primary analysis
Fig. 2
figure 2

Predictor importance plots: A Full algorithm and B Reduced algorithm. Abbreviations: BMI: Body Mass Index; P6M: In the past 6 months; CD4: Cluster of differentiation 4 receptor; EQ-5D-3L: EuroQoL-5Dimension-3Level; RNA: Ribonucleic acid; Hep B: Hepatitis B virus

The results for the additional analyses A-D are shown in Table 4 and summarized here. A) When using information from a single visit per individual, the overall performance was much lower, with AUC of 0.74 vs 0.82 for the full algorithm and 0.60 vs 0.76 for the reduced algorithm. B) For algorithms with the additional CES-D-10 thresholds, the AUC point estimate was higher for the full algorithm for cut-off 15 compared to 10 (0.87 vs 0.82). The other cut-off estimates were similar for to the corresponding algorithms for cut-off 10, with overlapping confidence intervals. C.) The algorithm including the additional predictor of food insecurity had a similar AUC estimate to the full algorithm, with overlapping confidence intervals. D.) The full algorithm predicting continuous CES-D-10 scores had a R-squared of 0.5 indicating that the algorithm explained only 50% of the variability in the CES-D-10 scores and had a high RMSE of 4.8, while for the reduced algorithm with a r-squared of 0.3, the algorithm explained only 30% of the variability in the scores and also had a high RMSE of 5.5.

Table 4 Comparison of performance evaluation measures for primary and additional analyses

Discussion

We developed a random forest algorithms using patient data from a cohort study that reliably predicted the presence of depressive symptoms indicative of a risk of clinical depression in a vulnerable HIV-HCV co-infected population. The algorithms used a set of selected candidate predictors (x = 137) from the cohort. The full algorithm using all candidate predictors performed better, with an AUC of 0.82, which indicates a 82% chance of distinguishing between CES-D-10 classes compared 0.76 for the reduced algorithm which used a smaller subset (x = 46) [33]. The prevalence of depressive symptoms was very high in our study, with more than 50% individuals found to be at risk for depression by CES-D-10 at their first visit. Despite this, only 10% had a documented depression diagnosis in their medical record suggesting there could be a substantial underestimation of the burden of depressive illness in this population without screening.

We developed this tool to assess if patient data that is commonly collected in clinical charts and research studies could be useful in predicting presence of depressive symptoms, which is seldom directly measured routinely for all patients, nor measured repeatedly over time. This tool will be most useful for conducting longitudinal clinical and epidemiologic research rather than for clinical care. It may prove useful to help identify people at risk for depression, study how this risk changes over time and with various interventions.

Most studies using machine learning have predicted the future onset of depressive symptoms [19, 40, 41] while a few, like ours, have focused on current depression prediction [42, 43]. A variety of predictors including demographic and clinical data, past medical history and life events have been studied. A range of machine learning algorithms like artificial neural networks, support vector machines, naïve Bayes classifier, and random forest were used in the general population and for specific groups like geriatric population and people with diabetes. These algorithms yielded AUC measures similar to ours, ranging between 0.70–0.95.

CCC collects extensive demographic, behavioral and clinical data. Using the full range and diversity of available predictors did show excellent discrimination in the full algorithm. However, for greater applicability, we chose to use a subset of 46 predictors that may be more readily available in other research settings, and despite using one third the number of predictors and less granular data, the overall discrimination was still acceptable. The additional analysis using only one visit per participant had a comparatively lower AUC, which may have been due to the smaller sample size and thus lower variability in the available data.

The algorithm we developed was a purely prediction algorithm and hence estimation of the strength of the effect of individual predictors is not possible. Further analysis with different modeling strategies would be needed for this assessment. However, the algorithm does provide some insight into factors that may be important for classification. The five most important predictors are related two main themes—i. SES (education, revenue source and employment) and ii. overall health status (HIV clinical stage and BMI). SES is a known strong determinant of depression. Receiving welfare and being from a low-income household has been associated with an elevated risk of food insecurity, and mental health issues [44, 45]. In Canada, almost 20% of people with major depression have been reported to be unemployed [46]. Another important health status related predictor was BMI. There have been studies with conflicting results regarding association between BMI and depression, and possible difference across race and gender [47, 48] and that BMI categories may not adequately capture people’s health status and thus this predictor needs to be considered with caution [49]. Finally, in the full algorithm with all 137 predictors, the EQ-5D-3L anxiety/depression dimension was the most important predictor and all EQ-5D dimensions (mobility, self-care, usual activities, pain/discomfort, and health state) were among the 25 most important. This provides further evidence that participant’s health status, and in the case of EQ-5D-3L, their perceived health status, are important in predicting depressive symptoms.

This study thus has many strengths. The CCC is generalizable to the HIV-HCV co-infected patients engaged in care in Canada, due to the recruitment from a variety of clinical settings (outreach, primary and tertiary care clinics in urban and semi-urban areas across the country). The sample used to develop these algorithms was generalizable to the parent CCC (see Appendix D; Table S3). The methodology used, RF, is non-parametric, highly accurate, and relatively robust to outliers, noise and does have safeguards from overfitting and thus improves chances of applicability beyond the data. Nevertheless, external validation is needed before application in other cohorts and research, to mitigate the risk of overfitting. In addition, the predictor importance plots provided some insight regarding predictors that play a major role in the accurate prediction of depressive symptoms.

The study however has limitations. The sample size is small as compared to big data applications of RF using electronic health records. Some predictors described in other studies such as childhood trauma, food insecurity among others, were not available for the full CCC. For example, in the additional analysis where we added the food security variable that was collected only in the food security sub-study was included, the AUC was slightly higher. Additionally, we categorized the CES-D-10 to create the binary classes, and thus may have lost some data by not predicting the individual CES-D-10 scores. We did develop a regression algorithm to predict the continuous CES-D-10 scores in additional analysis E, but it could only explain a small portion of the variability in the outcome. The gold standard depression diagnosis was not available in this study and thus the validity of the cut-off of 10 could not be assessed directly in this sample. In general, the overall AUCs were similar when using three other suggested CES-D- 10 cut-offs (8, 13, and 15) compared to a cut-off of 10. However, the full algorithm using a cut-off of 15 appears to have a higher AUC (0.87) than that of using a cut-off 10 (0.82). It will be important to assess in future studies whether this higher threshold may be more applicable to the co-infected population. However, since the CES-D-10 cut-off of 10 has been validated in HIV populations in Canada [24], we decided to use this threshold for comparability with available literature and future studies which may use this common threshold.

With a high proportion of participants with depressive symptoms in this population, it is important not to miss possible cases. Even if the algorithms we developed are considered to have acceptable discrimination (≥ 0.7) based on arbitrary thresholds, we would still misclassify a fair proportion of cases and thus this possible misclassification needs to be considered. Finally, this algorithm is applicable when the majority of the predictors are collected. However, in settings where such data is not available, especially completely clinical non-research setting implementing routine screening tools like the CES-D-10 should be considered, especially given the high prevalence of depressive symptoms we observed in this co-infected population.

Conclusions

Depressive symptoms indicative of a risk for clinical depression were common in our population of people living with HIV-HCV co-infection. The random forest algorithms we developed shows promise in accurately predicting an elevated risk of clinical depression using data on patient characteristics collected in research settings. The algorithms identified important characteristics for depressive symptoms classification including employment, HIV clinical stage, revenue source, BMI, and education. Such machine learning algorithms can be used in research settings especially cohort studies where such data may be available to predict presence of depressive symptoms and use this information to understand the impact of depressive symptoms on clinical, health service and patient-reported outcomes in vulnerable populations.

Availability of data and materials

The datasets generated and/or analysed during the current study are not publicly available. According to the stipulations of patient consent provided and our Institutional Ethics review boards, study records including confidential information collected during the study must be stored securely for 25 years after study completion, as required by Canadian clinical trial regulations. However, data stripped of personal identifiers, may be shared upon request to the corresponding author, or to: Mr. Sheldon Levy, Clinical Trials 2 Research Ethics Board (REB) Coordinator, MUHC Centre for Applied Ethics (sheldon.levy@muhc.mcgill.ca).

Abbreviations

AIDS:

Acquired Immunodeficiency Syndrome

APRI:

Aspartate Aminotransferase (AST) to platelet ratio

AUC:

Area under the Receiver Operating Characteristic Curve

BMI:

Body Mass Index

CCC:

Canadian Co-infection Cohort

CD4:

Cluster of differentiation 4 receptor

CDC:

Centre for Disease Control

CES-D-10:

Center for Epidemiologic Studies Depression Scale-10

CI:

Confidence Interval

EQ-5D-3L:

EuroQol-5 Dimension-3 Level

FS:

Food Security

HCV:

Hepatitis C virus

Hep B:

Hepatitis B virus

HFSSM:

Household Food Security Survey Module

HIV:

Human Immunodeficiency Virus

HR-QoL:

Health-related Quality of Life

IQR:

Interquartile Range

LR-:

Negative likelihood ratio

LR + :

Positive likelihood ratio

NPV:

Negative Predictive Value (NPV),

OOB:

Out-of-Bag Samples

P6M:

In the past 6 months

PPV:

Positive Predictive Value

RF:

Random Forests

RMSE:

Root Mean Squared Error

RNA:

Ribonucleic acid

ROC:

Receiver Operating Characteristic

SES:

Socioeconomic Status

References

  1. Shatte ABR, Hutchinson DM, Teague SJ. Machine learning in mental health: a scoping review of methods and applications. Psychol Med. 2019;49(9):1426–48.

    PubMed  Article  Google Scholar 

  2. Guidelines for the care and treatment of persons diagnosed with chronic hepatitis C virus infection. Geneva: World Health Organization; 2018. Report No.: License: CC BY-NC-SA 3.0 IGO. https://www.who.int/publications/i/item/9789241550345.

  3. Nanni MG, Caruso R, Mitchell AJ, Meggiolaro E, Grassi L. Depression in HIV infected patients: a review. Curr Psychiatry Rep. 2015;17(1):530.

    PubMed  Article  Google Scholar 

  4. Younossi Z, Park H, Henry L, Adeyemi A, Stepanova M. Extrahepatic manifestations of hepatitis C: a meta-analysis of prevalence, quality of life, and economic burden. Gastroenterology. 2016;150(7):1599–608.

    PubMed  Article  Google Scholar 

  5. Yeoh SW, Holmes ACN, Saling MM, Everall IP, Nicoll AJ. Depression, fatigue and neurocognitive deficits in chronic hepatitis C. Hepatol Int. 2018;12(4)294–304.

  6. Fialho R, Pereira M, Rusted J, Whale R. Depression in HIV and HCV co-infected patients: a systematic review and meta-analysis. Psychol Health Med. 2017;22(9):1089–104.

    PubMed  Article  Google Scholar 

  7. Belenky NM, Cole SR, Pence BW, Itemba D, Maro V, Whetten K. Depressive symptoms, HIV medication adherence, and HIV clinical outcomes in Tanzania: a prospective, observational study. PLoS ONE. 2014;9(5):e95469.

    PubMed  PubMed Central  Article  Google Scholar 

  8. Aibibula W, Cox J, Hamelin A-M, Moodie EEM, Anema A, Klein Marina B, et al. Association between depressive symptoms, CD4 count and HIV viral suppression among HIV-HCV co-infected people. AIDS Care. 2018;30(5):643–9.

    PubMed  Article  Google Scholar 

  9. Malhi GS, Mann JJ. Depression. Lancet (London, England). 2018;392(10161):2299–312.

    Article  Google Scholar 

  10. Eaton WW. Johns Hopkins Bloomberg School of Public Health Department of Mental H. Public mental health. New York: Oxford University Press; 2012.

    Google Scholar 

  11. Ma H, Villalobos CF, St-Jean M, Eyawo O, Lavergne MR, Ti L, et al. The impact of HCV co-infection status on healthcare-related utilization among people living with HIV in British Columbia, Canada: a retrospective cohort study. BMC Health Serv Res. 2018;18(1):319.

    PubMed  PubMed Central  Article  Google Scholar 

  12. Grebely J, Oser M, Taylor LE, Dore GJ. Breaking down the barriers to Hepatitis C Virus (HCV) treatment among individuals with HCV/HIV coinfection: action required at the system, provider, and patient levels. J Infect Dis. 2013;207(1):S19–25.

    PubMed  PubMed Central  Article  Google Scholar 

  13. Bonner JE, Barritt AST, Fried MW, Evon DM. Time to rethink antiviral treatment for hepatitis C in patients with coexisting mental health/substance abuse issues. Dig Dis Sci. 2012;57(6):1469–74.

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  14. Knott A, Dieperink E, Willenbring ML, Heit S, Durfee JM, Wingert M, et al. Integrated psychiatric/medical care in a chronic hepatitis C clinic: effect on antiviral treatment evaluation and outcomes. Am J Gastroenterol. 2006;101(10):2254–62.

    PubMed  Article  Google Scholar 

  15. Anagnostopoulos A, Ledergerber B, Jaccard R, Shaw SA, Stoeckle M, Bernasconi E, et al. Frequency of and risk factors for depression among participants in the Swiss HIV Cohort Study (SHCS). PLoS ONE. 2015;10(10):e0140943.

    PubMed  PubMed Central  Article  Google Scholar 

  16. Dwyer DB, Falkai P, Koutsouleris N. Machine learning approaches for clinical psychology and psychiatry. Annu Rev Clin Psychol. 2018;14(1):91–118.

    PubMed  Article  Google Scholar 

  17. Graham S, Depp C, Lee EE, Nebeker C, Tu X, Kim HC, et al. Artificial intelligence for mental health and mental illnesses: an overview. Curr Psychiatry Rep. 2019;21(11):116.

    PubMed  PubMed Central  Article  Google Scholar 

  18. Sau A, Bhakta I. Artificial Neural Network (ANN) model to predict depression among geriatric population at a Slum in Kolkata. India J Clin Diagn Res. 2017;11(5):Vc01-vc4.

    PubMed  Google Scholar 

  19. Jin H, Wu S, Di Capua P. Development of a clinical forecasting model to predict comorbid depression among diabetes patients and an application in depression screening policy making. Prev Chronic Dis. 2015;12:E142.

    PubMed  PubMed Central  Google Scholar 

  20. Klein MB, Saeed S, Yang H, Cohen J, Conway B, Cooper C, et al. Cohort profile: the Canadian HIV-hepatitis C co-infection cohort study. Int J Epidemiol. 2010;39(5):1162–9.

    PubMed  Article  Google Scholar 

  21. Cox J, Hamelin AM, McLinden T, Moodie EE, Anema A, Rollet-Kurhajec KC, et al. Food insecurity in HIV-Hepatitis C Virus Co-infected Individuals in Canada: the importance of co-morbidities. AIDS Behav. 2017;21(3):792–802.

    PubMed  Article  Google Scholar 

  22. EuroQol--a new facility for the measurement of health-related quality of life. Health Policy. 1990;16(3):199–208. https://doi.org/10.1016/0168-8510(90)90421-9.

  23. Andresen EM, Malmgren JA, Carter WB, Patrick DL. Screening for depression in well older adults: evaluation of a short form of the CES-D (Center for Epidemiologic Studies Depression Scale). Am J Prev Med. 1994;10(2):77–84.

    CAS  PubMed  Article  Google Scholar 

  24. Zhang W, O’Brien N, Forrest JI, Salters KA, Patterson TL, Montaner JS, et al. Validating a shortened depression scale (10 item CES-D) among HIV-positive people in British Columbia, Canada. PLoS ONE. 2012;7(7):e40793.

    CAS  PubMed  PubMed Central  Article  Google Scholar 

  25. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32.

    Article  Google Scholar 

  26. Malley JD, Kruppa J, Dasgupta A, Malley KG, Ziegler A. Probability machines: consistent probability estimation using nonparametric learning machines. Methods Inf Med. 2012;51(1):74–81.

    CAS  PubMed  Article  Google Scholar 

  27. Hastie T, Tibshirani R, Friedman JH. The elements of statistical learning : data mining, inference, and prediction. 2nd ed. New York: Springer; 2009. https://doi.org/10.1007/978-0-387-84858-7.

  28. Max Kuhn Contributions from Jed Wing SW, Andre Williams, Chris Keefer, Allan Engelhardt, Tony Cooper, Zachary Mayer and Brenton Kenkel, R Core Team, Michael Benesty, Reynald Lescarbeau, Andrew Ziem, Luca Scrucca, Yuan Tang, Can Candan, Tyler Hunt. caret: Classification and Regression Training. R package version 6.0–842019.

  29. Baron EC, Davies T, Lund C. Validation of the 10-item Centre for Epidemiological Studies depression scale (CES-D-10) in Zulu, Xhosa and Afrikaans populations in South Africa. BMC Psychiatry. 2017;17(1):6.

    PubMed  PubMed Central  Article  Google Scholar 

  30. Bjorgvinsson T, Kertz SJ, Bigda-Peyton JS, McCoy KL, Aderka IM. Psychometric properties of the CES-D-10 in a psychiatric sample. Assessment. 2013;20(4):429–36.

    PubMed  Article  Google Scholar 

  31. Canadian Community Health Survey, Cycle 2.2, Nutrition (2004): Income-Related Household Food Security in Canada. Ottawa: Health Canada; 2007. ISBN 978-0-662-45595-0. https://www.canada.ca/en/health-canada/services/food-nutrition/food-nutrition-surveillance/health-nutrition-surveys/canadian-community-health-survey-cchs/canadian-community-health-survey-cycle-2-2-nutrition-2004-income-related-household-food-security-canada-health-canada-2007.html.

  32. Steyerberg EW. Clinical prediction models : a practical approach to development, validation, and updating. Cham, Switzerland: Springer; 2019.

    Book  Google Scholar 

  33. Hosmer DW, Lemeshow S, Sturdivant RX. Applied Logistic Regression. Chicester: Wiley; 2013.

    Book  Google Scholar 

  34. Youden WJ. Index for rating diagnostic tests. Cancer. 1950;3(1):32–5.

    CAS  PubMed  Article  Google Scholar 

  35. StataCorp. Stata Statistical Software: Release 16. College Station, TX: StataCorp LLC; 2019.

  36. R Core Team. R: A Language and Environment for Statistical Computing. Vienna, Austria2013.

  37. Wright MN, Ziegler A. ranger: A Fast Implementation of Random Forests for High Dimensional Data in C++ and R. 2017. 2017;77(1):17.

  38. Wong J, Manderson T, Abrahamowicz M, Buckeridge DL, Tamblyn R. Can hyperparameter tuning improve the performance of a super learner?: a case study. Epidemiology. 2019;30(4):521–31.

    PubMed  PubMed Central  Article  Google Scholar 

  39. Joie E, Kym IES, Emma CM. PMCALPLOT: Stata module to produce calibration plot of prediction model performance. S458486 ed: Boston College Department of Economics; 2018.

  40. Sau A, Bhakta I. Screening of anxiety and depression among the seafarers using machine learning technology. Inform Med Unlocked. 2018:100149. https://doi.org/10.1016/j.imu.2018.12.004.

  41. Rosellini AJ, Liu S, Anderson GN, Sbi S, Tung ES, Knyazhanskaya E. Developing algorithms to predict adult onset internalizing disorders: An ensemble learning approach. J Psychiatr Res. 2019;121:189–96.

    PubMed  PubMed Central  Article  Google Scholar 

  42. Wang J, Sareen J, Patten S, Bolton J, Schmitz N, Birney A. A prediction algorithm for first onset of major depression in the general population: development and validation. J Epidemiol Community Health. 2014;68(5):418–24.

    PubMed  Article  Google Scholar 

  43. King M, Bottomley C, Bellon-Saameno J, Torres-Gonzalez F, Svab I, Rotar D, et al. Predicting onset of major depression in general practice attendees in Europe: extending the application of the predictD risk algorithm from 12 to 24 months. Psychol Med. 2013;43(9):1929–39.

    CAS  PubMed  Article  Google Scholar 

  44. Coiro MJ. Depressive symptoms among women receiving welfare. Women Health. 2001;32(1–2):1–23.

    CAS  PubMed  Article  Google Scholar 

  45. Wu S, Fraser MW, Chapman MV, Gao Q, Huang J, Chowa GA. Exploring the relationship between welfare participation in childhood and depression in adulthood in the United States. Soc Sci Res. 2018;76:12–22.

    PubMed  PubMed Central  Article  Google Scholar 

  46. Rizvi SJ, Cyriac A, Grima E, Tan M, Lin P, Gallaugher LA, et al. Depression and employment status in primary and tertiary care settings. Can J Psychiatry. 2015;60(1):14–22.

    PubMed  PubMed Central  Article  Google Scholar 

  47. Monda V, La Marra M, Perrella R, Caviglia G, Iavarone A, Chieffi S, et al. Obesity and brain illness: from cognitive and psychological evidences to obesity paradox. Diabetes Metab Syndr Obes. 2017;10:473–9.

    PubMed  PubMed Central  Article  Google Scholar 

  48. Banack HR, Kaufman JS. From bad to worse: collider stratification amplifies confounding bias in the “obesity paradox.” Eur J Epidemiol. 2015;30(10):1111–4.

    PubMed  Article  Google Scholar 

  49. Nuttall FQ. Body mass index: obesity, bmi, and health: a critical review. Nutr Today. 2015;50(3):117–28.

    PubMed  PubMed Central  Article  Google Scholar 

Download references

Acknowledgements

We would like to acknowledge the participants of the Canadian Co-Infection Cohort (CTN222), the study coordinators and nurses for their assistance with study coordination, participant recruitment, and care, and the Canadian Co-Infection Cohort (CTN222) co-investigators—Drs. Lisa Barrett, Jeff Cohen, Brian Conway, Curtis Cooper, Pierre Côté, Joseph Cox, M. John Gill, Shariq Haider, David Haase, Mark Hull, Valérie Martel-Laferrière, Julio Montaner, Erica E. M. Moodie, Neora Pick, Danielle Rouleau, Aida Sadr, Steve Sanche, Roger Sandre, Mark Tyndall, Marie-Louise Vachon, Sharon Walmsley and Alexander Wong.

Consortium name: Canadian Co-infection Cohort

Principal investigator: Marina B. Klein1, 2, 11

Co-investigators: Lisa Barrett12, Jeff Cohen13, Brian Conway5, Curtis Cooper4, Pierre Côté14, Joseph Cox1, 2, John Gill15, Shariq Haider16, Mark Hull6, Valérie Martel-Laferrière7, Erica E. M. Moodie1, Neora Pick17, Danielle Rouleau18, Steve Sanche19, Roger Sandre20, Marie-Louise Vachon8, Sharon Walmsley9, Alexander Wong10

12. Dalhousie University, Halifax, Nova Scotia, Canada

13. Windsor Regional Hospital Metropolitan Campus, Windsor, Ontario, Canada

14. Clinique Médicale du Quartier Latin, Montreal, Quebec, Canada

15. Southern Alberta HIV Clinic, Calgary, Alberta, Canada

16. McMaster University, Hamilton, Ontario, Canada

17. Oak Tree Clinic, Vancouver, British Columbia, Canada

18. Université de Montréal, Montreal, Quebec, Canada

19. University of Saskatchewan, Saskatoon, Saskatchewan, Canada

20. Sudbury Regional Hospital, Sudbury, Ontario, Canada

Funding

This work was supported by Fonds de recherche du Québec-Santé; Réseau sida/maladies infectieuses, the Canadian Institute for Health Research (CIHR; FDN-143270); and the CIHR Canadian HIV Trials Network (CTN222 & CTN264). GM is supported by the PhD trainee fellowship from the Canadian Network on Hepatitis C. MBK is supported by a Tier I Canada Research Chair. The funders had no role in the production of this manuscript. EEMM is supported by a chercheur de mérite award from the Fonds de recherche du Québec-Santé and a Canada Research Chair (Tier 1). VML is supported by Clinical Research Scholars–Junior 1 from the Fonds de recherche du Québec-Santé.

Author information

Authors and Affiliations

Authors

Consortia

Contributions

All authors contributed to this study, as required by the International Committee of Medical Journal Editors. MBK, EEMM and GM conceived of and designed the study. GM and CLD prepared the analytical dataset. GM performed all statistical analyses. GM, EEMM and MBK drafted the initial manuscript. All co-authors MJB, JC, CC, BC, MH, VML, MLV, SW, and AW revised the document critically and gave final approval prior to completion. All authors take responsibility for the accuracy and integrity of this work. The author(s) read and approved the final manuscript.

Authors’ information

Not applicable.

Corresponding author

Correspondence to Marina B. Klein.

Ethics declarations

Ethics approval and consent to participate

This study was approved by the Research Ethics Board of the McGill University Health Centre (2021–6985). The CCC and the FS Sub-Study were approved by the Research Ethics Board of the McGill University Health Centre (2006–1875, BMB-06-006t, 2013–994) and the research ethics boards of participating institutions. The study was conducted according to the Declaration of Helsinki. Informed consent was obtained from all individual participants included in the study.

Consent for publication

Not applicable.

Competing interests

JC received grants and consulting fees from ViiV Healthcare, Merck, and Gilead and personal fees from Bristol-Myers Squibb. CC has received personal fees for being a member of the national advisory boards of Gilead, Merck, Janssen, and Bristol-Myers Squibb. BC is a board member, consultant, and has received grants and payment for lectures from AbbVie, Gilead, and Merck, and payment for educational presentations from AbbVie. MH has served as a consultant for Merck, Vertex Pharmaceuticals, Pfizer, Viiv Healthcare, and Ortho-Jansen. MH has also received grants from the National Institute on Drug Abuse, as well as payment for lectures from Merck and Ortho-Janssen. MLV reports personal fees from Abbvie, personal fees from Merck, personal fees from Gilead, outside the submitted work. SW received grants, consulting fees, lecture fees, nonfinancial support, and fees for the development of educational presentations from Merck, ViiV Healthcare, GlaxoSmithKline, Pfizer, Gilead, AbbVie, Bristol-Myers Squibb, and Janssen. MBK reports grants for investigator-initiated studies from ViiV Healthcare, AbbVie, Merck, and Gilead; and consulting fees from ViiV Healthcare, Merck, AbbVie, and Gilead. GM, EEMM, MJB, CLD, VML, and AW have no conflicts of interest to disclose.

Additional information

Publisher’s Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Marathe, G., Moodie, E.E.M., Brouillette, MJ. et al. Predicting the presence of depressive symptoms in the HIV-HCV co-infected population in Canada using supervised machine learning. BMC Med Res Methodol 22, 223 (2022). https://doi.org/10.1186/s12874-022-01700-y

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s12874-022-01700-y

Keywords

  • HIV-HCV co-infection
  • Depressive symptoms
  • Supervised machine learning
  • Random forests