Integrated youth services, often abbreviated as IYS, represent a growing model of care in which young people can access a range of youth-friendly health and social services within a single location. Rather than forcing adolescents and young adults to navigate fragmented primary care, mental health clinics, substance use programs, and social services on their own, IYS centres consolidate these supports under one roof. Foundry, located in British Columbia, was the first such initiative established in Canada, and it has since become a testing ground for understanding how this model performs in practice. The new study, led by Katherine G. Hastings of the University of British Columbia’s School of Population and Public Health, together with colleagues across UBC, Foundry, and Providence Research, examined service utilization patterns and characteristics of youth reporting suicidal ideation to better understand and address their care needs.

The methodological backbone of the study is a retrospective analysis of linked health survey and service data collected at Foundry from 2018 to 2023. The researchers compared two groups: youth who reported having suicidal ideation within the past month at their initial visit, and youth who reported never having suicidal ideation at that time. To quantify differences between these groups, the team calculated standardized mean differences, or SMDs, alongside conventional p-values. Standardized mean differences are a statistical tool that expresses the gap between two groups in units of pooled standard deviation, which makes them particularly useful for comparing variables measured on different scales and for detecting imbalances in cohort studies. Values above roughly 10 percent, expressed here in the study’s convention as SMDs in the tens, typically signal a meaningful difference between groups, independent of sample size.

The results paint a stark picture of inequity. Youth with past-month suicidal ideation were substantially more likely to identify as gender diverse, at 13.3 percent compared with 4.9 percent among youth reporting no suicidal ideation, yielding a standardized mean difference of 29.7. They were also far more likely to identify as LGBTQ+, at 45.2 percent versus 22.3 percent, an SMD of 52.7. These are not marginal statistical artifacts; they represent a doubling or tripling of representation relative to their peers without recent suicidal thoughts. The findings align with a broader body of evidence that gender and sexually diverse youth face elevated risks of psychological distress and suicidality, and they underscore that integrated youth services are serving a population in which sexual and gender minority youth are heavily represented.

Perhaps the most striking quantitative signal in the entire study concerns psychological distress. Youth with past-month suicidal ideation scored 36.7 on the Kessler Psychological Distress Scale, a widely used ten-item screening instrument that measures symptoms of anxiety and depression over the preceding month, compared with a mean score of 27.0 among youth who reported never having suicidal ideation. The standardized mean difference of 138.5 is enormous by any epidemiological standard, indicating that the two groups occupy almost entirely different regions of the distress distribution. For clinicians and service planners, this suggests that the youth presenting with recent suicidal thoughts at Foundry are not merely experiencing transient worry but profound, clinically significant psychological suffering that demands immediate and intensive response.

Service utilization data add another layer of nuance. Counterintuitively, youth with past-month suicidal ideation had fewer contacts with the integrated youth services initiative than those without suicidal ideation, averaging 7.0 visits with a standard deviation of 16.8, compared with 11.2 visits with a standard deviation of 23.2 among youth without suicidal ideation, a difference that was statistically significant at p less than 0.0001. The large standard deviations reveal enormous variability in how many times individual youth return, with some attending dozens of visits and others only one or two. At the same time, youth with recent suicidal ideation were more likely to have accessed other mental health or substance use services in the preceding year, at 43.4 percent versus 20.9 percent, an SMD of 49.7. Taken together, these patterns suggest that many youth with recent suicidal thoughts arrive at Foundry already known to other parts of the health system, yet their engagement with the integrated model itself may be less sustained, raising questions about whether the model is fully retaining the youth who need it most.

Tracking the monthly proportion of youth reporting suicidal ideation across the five-year study period, the researchers found no substantial change overall. This temporal stability is itself informative: it suggests that the burden of suicidal thinking among youth seeking integrated services neither worsened dramatically nor eased during a window that included the COVID-19 pandemic and its aftermath, a period widely associated with heightened strain on youth mental health. However, when the team disaggregated the trends by subpopulations, they observed a slight increase among females in the last two years of the study. This sex-specific uptick, though modest, echoes concerns raised by clinicians and public health officials in multiple countries about deteriorating mental health indicators among adolescent girls and young women, and it signals a group that may warrant targeted monitoring and intervention.

The study’s implications extend beyond British Columbia. Suicidal ideation in adolescence and young adulthood is one of the strongest known predictors of subsequent suicide attempts, and early identification is a cornerstone of suicide prevention strategy. The fact that nearly three-quarters of youth accessing Foundry reported some lifetime history of suicidal thoughts, and more than a third reported thoughts within the past month, transforms these centres into critical detection sites. As the authors conclude, the sheer prevalence of suicidal ideation among service users indicates an important opportunity for early intervention. A young person who comes to an integrated youth centre for what seems like a routine concern may, on systematic screening, disclose recent suicidal thinking, and that moment represents a chance to intervene before ideation progresses to action.

The findings also shed light on health inequities in self-reported suicidal ideation, particularly among gender and sexually diverse youth. The overrepresentation of LGBTQ+ and gender-diverse youth among those with recent suicidal thoughts is consistent with minority stress frameworks, which attribute elevated mental health risk among sexual and gender minorities to experiences of stigma, discrimination, and social exclusion rather than to any intrinsic vulnerability. For integrated youth services, this means that culturally safe, affirming care is not an optional enhancement but a core requirement. Services that fail to create welcoming environments for these youth risk losing contact with precisely the group at highest risk, and the data on lower visit counts among youth with recent ideation hint that engagement itself may be fragile.

Methodologically, the study has the strengths and limitations inherent to its design. As a retrospective cohort analysis of youth who voluntarily accessed a single integrated youth services initiative, it captures the population that reaches care, not the broader population of youth with suicidal ideation who never seek help. Survey responses at the initial visit are self-reported and may be shaped by willingness to disclose sensitive information, although the voluntary consent process and the youth-centred culture of Foundry likely support honest reporting. The research ethics approval was obtained by the University of British Columbia and Providence Health Care Research Ethics Board, and participants consented to research use of their data. Notably, the authors acknowledge the contribution of Foundry’s youth and family engagement team, who reviewed the study and assisted with co-interpretation of the results, an approach that reflects the participatory ethos of the integrated youth services movement itself. The work was funded by the Canadian Institutes of Health Research, with the funder playing no role in study design, analysis, interpretation, or writing. As integrated youth services continue to expand across Canada and beyond, this study provides a rigorous baseline for understanding who these services reach, how deeply their users are suffering, and where the greatest opportunities for suicide prevention lie.

Subject of Research: Suicidal ideation and service utilization among youth accessing integrated youth services in British Columbia

Article Title: Characteristics of youth reporting suicidal ideation in an integrated youth services initiative in British Columbia, Canada (2018–2023): a retrospective cohort study

Article References: Hastings, K. G., Liu, G., Marchand, K., Gadermann, A., Sutherland, J. M., Jenkins, E., & Barbic, S. P. (2026). Characteristics of youth reporting suicidal ideation in an integrated youth services initiative in British Columbia, Canada (2018–2023): a retrospective cohort study. BMC Pediatrics, 26(1), Article 870. https://doi.org/10.1186/s12887-026-07543-6

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07543-6

Keywords: suicidal ideation, integrated youth services, Foundry, adolescent mental health, LGBTQ+ youth, health equity, psychological distress, suicide prevention, retrospective cohort study, British Columbia, service utilization, gender diversity

The research, conducted by Yutong Guo and Chao Yang, began with a problem familiar to every foundry engineer: pouring-process parameters have traditionally been set by experience and adjusted through repeated, costly iterations. When several parameters interact, operating conditions fluctuate, and multiple defect types appear simultaneously, this conventional approach suffers from slow adjustment cycles and difficulty pinpointing which factors actually drive defects. Industrial digitalization, however, has quietly accumulated vast archives of shop-floor records, and the authors recognized that this data could serve as the foundation for a closed-loop workflow: clean the data, predict quality, screen for likely defect causes, optimize the adjustable parameters, and then verify the recommendations in real production.

The dataset came from actual pouring records at a cooperating foundry producing large, complex castings. The raw template contained 65 fields: 49 input variables covering chemical composition, molding-sand properties, and temperature conditions; a product qualification rate; pouring and scrap quantities; and 13 categories of scrap-cause labels drawn from routine quality-inspection records. Of 354 raw records, four incomplete entries were removed, leaving 350 valid records for modeling. Because full-volume radiographic or industrial CT inspection was impractical for such large components, the factory relied on destructive sectioning at designated positions followed by visible-color liquid penetrant testing, in line with ISO 4987:2020 and ISO 3452-1:2021, exposing internal discontinuities such as sand holes and shrinkage porosity on the freshly cut surfaces.

With the data cleaned, the team compared four candidate regression models for predicting qualification rate: a feedforward neural network, k-nearest neighbors regression, Gaussian process regression, and a random forest. Under five-fold cross-validation, the 150-tree random forest achieved the lowest root mean square error, at 3.6865 percentage points, along with a mean absolute error of 2.5832. A sensitivity analysis varying the ensemble size from 10 to 500 trees confirmed that 150 was sufficient; larger forests produced no further error reduction. Random forest offered an additional practical advantage: it provides explicit feature-relevance rankings, which the researchers needed to identify which of the 49 variables deserved attention, and it could be evaluated rapidly thousands of times during a subsequent optimization search.

Feature relevance pointed to chemical-composition elements such as calcium, magnesium, carbon, sulfur, and silicon, along with AFS sand fineness and temperature-related variables. But the six variables ultimately selected for direct adjustment were not simply the six highest-ranked. The authors jointly weighed model relevance, shop-floor adjustability, and engineering feasibility, arriving at moisture content, hot-wet tensile strength, tapping temperature, AFS fineness, permeability, and sand temperature—factors that foundry crews can realistically control. Two-dimensional conditional-response surfaces showed that predicted qualification rate changed in interval-dependent, partially interacting ways across variable combinations, and Friedman H statistics quantified measurable but moderate non-additive interactions, the largest between tapping temperature and permeability.

The second computational stage tackled a subtler question: could routine scrap-cause records, rather than just the single qualification-rate number, inform the optimization? Because several defect causes can be recorded under the same process conditions, the team formulated the task as a multi-label screening problem rather than a mutually exclusive classification. A deep feedforward neural network with a 49–128–64–13 architecture—two ReLU-activated hidden layers feeding a 13-neuron sigmoid output—was trained on the 49 process parameters to produce scores for all 13 scrap-cause categories simultaneously. Given the extreme sparsity of the labels, only three categories had at least five positive records: sand hole, sand adhesion, and shrinkage porosity. Decision thresholds were calibrated on a validation subset instead of being fixed at 0.5, and the network was evaluated with standard multi-label metrics including Micro-F1, Hamming loss, and subset accuracy.

Crucially, the researchers did not treat the neural network as a second optimization objective. Instead, its label scores served as feasibility constraints during candidate screening: a proposed parameter setting was accepted only if the context-averaged scores for sand hole, sand adhesion, and shrinkage porosity did not increase relative to the historical baseline. This design kept the random forest as the primary evaluator of quality while the DNN acted as a guardrail against settings that might trade a higher predicted qualification rate for a higher risk of specific defects. It is an elegant division of labor—prediction, screening, and search each handled by a model suited to its role, all operating on the same production context.

The optimization itself used an evolutionary search implemented in the DEAP framework, with a population of 100, crossover probability of 0.7, and mutation probability of 0.3, searching within bounds set by observed historical values. A supplementary verification running the search for 100 generations across 10 independent runs showed that the best feasible solution at generation 2 was already within roughly 0.008 percentage points of the eventual plateau, evidence of rapid convergence. The random-forest surrogate estimated that replacing the six variables with optimized-center values would raise the context-averaged qualification rate from 97.9441 percent to 98.7799 percent—a model-based counterfactual gain of about 0.84 percentage points. Importantly, the authors present this as a model estimate supporting parameter recommendation, not a causal claim of field improvement.

Equally notable is how the recommendations were expressed. Rather than a single theoretical optimum—impractical in a factory where raw materials, equipment states, and ambient conditions constantly shift—the team reported bounded ranges for each variable. Moisture content was to decrease slightly from a historical mean of 3.028 percent toward 2.866 to 3.023 percent; hot-wet tensile strength was to increase from 4.90 kPa into the 5.51 to 5.94 kPa window; AFS fineness was to rise from 57.54 toward 58.30 to 63.45; and permeability was to fall from 172.4 into the 140.0 to 165.9 range. Tapping and sand temperatures changed only marginally. Read together, these coupled adjustments point toward a more stable mold condition: finer sand with lower permeability and moisture, but stronger hot-wet tensile resistance against erosion and scabbing at the mold–metal interface.

The final proof came from the foundry itself. A pilot trial of five products produced under the recommended settings yielded five qualified castings. Scale-up followed across four consecutive batches of 50 products each, with qualification rates of 98.0, 100.0, 98.0, and 98.0 percent—197 of 200 products overall, or 98.5 percent, compared with a historical aggregate baseline of 98.17 percent. Sectioned penetrant inspection added physical corroboration: before the adjustment, localized red penetrant bleed-out indications were visible near the junction region of exposed sections; after implementation, no comparable indication appeared in the inspected region. Combined with the lower modeled shrinkage-porosity label score, the evidence suggests genuinely improved local defect conditions. The study stands out precisely because so many data-driven manufacturing papers stop at offline modeling—here, prediction, multi-label defect screening, constrained evolutionary search, and staged field validation form one continuous loop, offering foundries a realistic template for turning accumulated production data into executable process windows.

Subject of Research: Data-driven quality prediction and process-parameter optimization for large complex castings using machine learning and evolutionary search

Article Title: Data-driven quality prediction and process optimization for large complex castings

Article References: Guo, Y., & Yang, C. (2026). Data-driven quality prediction and process optimization for large complex castings. Results in Engineering, 32, Article 113269. https://doi.org/10.1016/j.rineng.2026.113269

Image Credits: AI Generated

DOI: 10.1016/j.rineng.2026.113269

Keywords: casting, machine learning, random forest, deep neural network, multi-label classification, evolutionary optimization, process optimization, defect prediction, foundry, liquid penetrant testing, qualification rate, manufacturing