Introduction

Non-suicidal self-injury (NSSI) is defined as the direct and deliberate destruction of one’s own body tissue without suicidal intent and represents a substantial public health concern. It is particularly prevalent among adolescents and psychiatric populations, in whom it may function as a maladaptive response to overwhelming emotional distress.1,2 A recent meta-analysis estimated that the co-occurrence of NSSI and suicide attempts among individuals with mental disorders was approximately 26%.2 Recurrence is also common, with a 6-month recurrence rate of up to 65.19% reported among adolescents after psychiatric hospitalization.3 The frequency and severity of NSSI are associated with considerable psychological distress and an increased risk of subsequent suicide attempts, highlighting the need for timely risk assessment and preventive intervention.4

Accurate prediction of NSSI recurrence remains challenging. Existing assessments rely largely on clinical interviews and self-reported information, which may be affected by recall bias, under-reporting, shame, and social desirability.5 Many previous studies have also used cross-sectional designs or conventional prediction approaches that do not adequately account for the structure of inpatient follow-up. In psychiatric wards, discharge without recurrence precludes any subsequent in-hospital NSSI event and therefore constitutes a competing event rather than ordinary censoring. Treating discharged patients as simply censored in standard Kaplan–Meier analyses assumes that they remain at risk of an inpatient event and may consequently overestimate cumulative recurrence risk. A competing-risk framework is therefore more appropriate for estimating the probability of NSSI recurrence before discharge. This issue is particularly relevant because the transition from inpatient to post-discharge care is itself a clinically vulnerable period.6

Objective biomarkers may complement clinical assessment and improve risk characterization. Systemic inflammation has increasingly been implicated in the pathophysiology of NSSI. Accessible peripheral inflammatory indices, including the neutrophil-to-lymphocyte ratio (NLR), monocyte-to-lymphocyte ratio, and C-reactive protein, have been associated with NSSI and related behavioral outcomes.7,8 These associations may reflect neuroimmune dysregulation involving inflammatory pathways and prefrontal–limbic circuits relevant to emotion regulation and behavioral control.8,9 However, inflammatory markers are nonspecific and are unlikely to provide sufficient predictive information when considered in isolation.

Accordingly, prediction of in-hospital NSSI recurrence may benefit from combining objective inflammatory markers with established clinical history, symptom severity, and early behavioral information. Integrating these complementary domains within a competing-risk model may provide a more complete estimate of short-term recurrence risk than any single source of information.10 Therefore, this study aimed to identify clinical and inflammatory predictors of subsequent in-hospital NSSI recurrence and to develop and internally validate a competing-risk nomogram that accounts for discharge without recurrence as a competing event. We also examined potential non-linear associations between continuous predictors and recurrence risk.

MethodsStudy Design

This retrospective, single-center cohort study based on real-world clinical data was conducted to evaluate the competing risks of adverse outcomes during hospitalization among patients with NSSI. We consecutively screened patients admitted to the psychiatric and psychological wards of our Hospital. The diagnosis of NSSI was established utilizing electronic medical records (EMR) at discharge. The study protocol was reviewed and approved by the Institutional Review Board of Hebei Provincial Mental Health Center (Approval No. 202012). Because this retrospective study involved analysis of existing de-identified medical records and posed minimal risk to participants, the requirement for individual informed consent, including parental or legal guardian consent for participants younger than 18 years, was waived by the Institutional Review Board.

Participants

The study cohort was identified through a systematic query of the institutional EMR system for patients aged 12 to 40 years admitted to the psychiatric or psychological wards between January 1, 2021, and June 30, 2025. International Classification of Diseases, Tenth Revision (ICD-10) codes for intentional self-harm (X60–X84) were utilized as a high-sensitivity initial screening strategy to capture all potential self-harm events. Eligible participants were subsequently selected via a three-stage validation process: (1) automated identification by the hospital’s clinical data warehouse according to predefined criteria; (2) manual review of electronic medical records, including admission notes, nursing logs, and psychological reports, by two independent psychiatrists to achieve high-specificity exclusion of suicide attempts and confirm adherence to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) criteria for NSSI;11 and (3) final adjudication by a senior psychiatrist in cases of diagnostic or eligibility uncertainty, with discrepancies resolved through consensus.

Inclusion criteria comprised all of the following: (1) Age between 12 and 40 years at the time of admission; (2) Confirmed diagnosis of NSSI upon admission or a clearly documented history of NSSI within the 12 months prior to hospitalization, establishing a baseline risk cohort for in-hospital recurrence according to DSM-5 criteria;11 (3) Total hospitalization duration exceeding 24 hours to ensure a sufficient observation window for the competing risk analysis.

Exclusion criteria encompassed any of the following: (1) Missing data for key clinical variables; (2) Self-injurious behaviors prior to or during hospitalization clinically judged to possess explicit suicidal intent (ie, Suicide Attempt), corroborated by clinical consensus or a score indicating active suicidal intent on the Columbia-Suicide Severity Rating Scale (C-SSRS);12 (3) Presence of severe intellectual disability, acute psychotic episodes (eg, active schizophrenia with command hallucinations), or severe cognitive impairment that precluded accurate assessment of self-harm intent; (4) Interruption of the observation period due to clinical necessity requiring transfer to an Intensive Care Unit (ICU) or other non-psychiatric medical wards.

Data Collection and Outcome Definition

Using the aforementioned dual-review process, clinical data were extracted from all EMR and laboratory information systems (LIS), including physician progress notes, nursing logs, psychiatric evaluations, and laboratory reports. Admission demographic and clinical characteristics and laboratory measurements obtained within the first 24 hours were recorded as baseline predictors. NSSI occurring after admission but within the first 24 hours was recorded separately as an early in-hospital behavioral variable rather than as an admission baseline characteristic.

The primary outcome was in-hospital NSSI recurrence, defined as any new intentional, non-suicidal self-injurious behavior documented after the first 24 hours of admission and before discharge. NSSI occurring within the first 24 hours was not counted as an outcome event but was treated as an early in-hospital behavioral predictor. Among patients with an NSSI event during this initial period, recurrence therefore referred to an additional event occurring after 24 hours. To account for the mutually exclusive nature of events in the inpatient setting, discharge without an NSSI event or in-hospital mortality was defined as a competing risk event, given that these occurrences inherently preclude the observation of subsequent inpatient self-harm. Since in-hospital mortality is exceedingly rare in this specific psychiatric cohort, the predominant competing event was survival discharge. The observation period commenced 24 hours post-admission and terminated at the time of the first occurrence of NSSI recurrence, discharge, or death. The utilization of a competing risk framework is methodologically crucial in longitudinal observational studies to avoid the overestimation of cumulative incidence that typically occurs with traditional Kaplan-Meier analyses.13

A comprehensive set of candidate predictor variables was systematically curated based on clinical relevance and existing literature regarding NSSI pathogenesis. Demographic data included age, sex, marital status, occupational status, educational level, and body mass index (BMI). Clinical baseline characteristics comprised primary psychiatric diagnoses (eg, bipolar disorder, major depressive disorder, borderline personality disorder), lifetime history of prior suicide attempts, age of NSSI onset, NSSI frequency within the preceding year, specific methods of self-harm, family history of psychiatric disorders, and documented lifetime history of childhood trauma or abuse. Somatic comorbidities were quantified using the Charlson Comorbidity Index (CCI).14 The baseline severity of psychiatric symptoms was assessed using the Clinical Global Impressions (CGI) scale,15 the Hamilton Depression Rating Scale (HAMD),16 and the Hamilton Anxiety Rating Scale (HAMA),17 derived from scores documented in routine clinical practice according to the institution’s standardized psychiatric admission assessment protocol. Baseline laboratory data, including CBC, inflammatory markers, and fasting metabolic indices, were restricted to the initial 24-hour admission window; from these, indices such as the NLR and PLR were calculated. Treatment-related variables detailed the specific psychotropic medication regimens (eg, mood stabilizers, antidepressants, second-generation antipsychotics) and psychotherapeutic treatment plans (eg, crisis intervention, brief supportive counseling) formulated within the first 24 hours of admission, focusing on acute-phase strategies rather than long-term structured therapies.

Statistical Analysis

All analyses were performed using R version 4.5.1. Variables with less than 20% missingness were imputed using multiple imputation by chained equations, and estimates were pooled according to Rubin’s rules. Continuous variables were assessed using the Shapiro–Wilk test and are presented as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables are reported as frequencies and percentages. Descriptive comparisons between the NSSI recurrence and competing-event groups were conducted using Student’s t-test or the Mann–Whitney U-test for continuous variables and the Pearson chi-square test or Fisher’s exact test for categorical variables. These comparisons were not used for predictor selection.

The cumulative incidence functions of NSSI recurrence and competing events were estimated, with discharge without recurrence and in-hospital death treated as competing events. Gray’s test was used to compare cumulative incidence curves between subgroups. Associations with NSSI recurrence were evaluated using Fine–Gray proportional subdistribution hazards models.

Candidate predictors for the multivariable model were prespecified according to clinical relevance, previous evidence, temporal availability, and routine measurability rather than significance in baseline comparisons or univariable analyses. The prespecified predictors were age, sex, primary borderline personality disorder diagnosis, lifetime history of prior suicide attempt, lifetime history of childhood trauma or abuse, use of multiple self-harm methods, NSSI during the first 24 hours, HAMD score, NLR, and crisis intervention during the first 24 hours. Age and sex were retained regardless of statistical significance. Univariable Fine–Gray analyses were reported descriptively and were not used for variable selection.

Multicollinearity was assessed using generalized variance inflation factors (GVIFs), calculated from the correlation matrix of the model coefficients. For comparability across terms with different degrees of freedom, GVIF(1/[2×df]) was reported. A value ≥2.24, corresponding approximately to a conventional VIF of 5 for a one-degree-of-freedom term, was considered indicative of substantial collinearity. When predictors represented overlapping psychiatric or inflammatory constructs, the variable with greater clinical relevance to the prespecified model was retained, and correlated predictors were not entered simultaneously.

Potential non-linear associations of continuous NLR and HAMD score with recurrence were examined using restricted cubic splines incorporated into the multivariable Fine–Gray model. Overall and non-linear associations were assessed using Wald tests. Continuous NLR constituted the primary analysis of its functional relationship with recurrence. Because no NLR threshold was prespecified, an exploratory cutpoint was identified using a maximally selected Gray statistic. Candidate cutpoints were restricted to observed NLR values between the 10th and 90th percentiles. At each candidate value, Gray’s statistic was calculated while retaining discharge without recurrence as a competing event. The value yielding the largest standardized statistic was selected, identifying an NLR cutpoint of 2.45. This analysis was implemented using the cuminc function in the cmprsk package (version 2.2–12). The dichotomized NLR variable was used only in the simplified nomogram and was considered exploratory.

The proportional subdistribution hazards assumption was assessed using time-varying coefficient tests. Each retained predictor was interacted with log-transformed follow-up time, and individual and joint time-dependent effects were evaluated using Wald tests. A p value <0.05 indicated evidence of non-proportionality. These analyses were implemented using the cov2 and tf arguments of the crr function.

All prespecified predictors without substantial collinearity were entered simultaneously into the primary multivariable Fine–Gray model, with continuous NLR modeled using restricted cubic splines. For construction of the simplified nomogram, a secondary Fine–Gray model was fitted using the same predictors but replacing the spline term for continuous NLR with the exploratory dichotomous variable NLR >2.45. Results are presented as subdistribution hazard ratios with 95% confidence intervals. Because NSSI during the first 24 hours represented an early manifestation of the same behavioral outcome, a sensitivity model excluding this variable was fitted, and the estimates of the remaining predictors and model C-index were compared with those of the primary model.

A competing-risk nomogram was constructed from predictors independently associated with recurrence to estimate 7- and 14-day risks. Internal validation was performed using 1000 bootstrap resamples with optimism correction. In each resample, the model-development procedure was repeated, including estimation of the exploratory NLR cutpoint and refitting of the Fine–Gray model. Model performance was calculated in the bootstrap sample and then evaluated in the original dataset. Optimism was defined as the average difference between these two estimates and was subtracted from the apparent performance obtained in the original cohort. Optimism-corrected discrimination was assessed using a competing-risk-adapted C-index. Calibration at 7 and 14 days was evaluated using calibration-in-the-large, calibration slope, and graphical comparison of predicted and observed cumulative incidence. Bias-corrected calibration curves were obtained by subtracting the mean bootstrap optimism from the apparent calibration estimates across the range of predicted risks. Decision curve analysis (DCA) was used to quantify clinical net benefit across threshold probabilities. All tests were two-sided, with p<0.05 considered statistically significant.

ResultsStudy Population and Baseline Characteristics

Of 1256 screened patients, 414 were excluded (missing data, n=112; explicit suicide attempts, n=268; non-psychiatric transfer, n=34), yielding a final analytical cohort of 842. Over a median observation period of 16.5 days (IQR, 9.0–24.0), 145 (17.2%) experienced in-hospital NSSI recurrence, 661 (78.5%) experienced the competing event (survival discharge), and 36 (4.3%) were censored. As detailed in Table 1, the overall cohort had a mean age of 21.4 ± 5.6 years and was predominantly female (77.2%). Descriptive comparisons were conducted between the NSSI recurrence and competing event groups. Compared to the competing event group, the NSSI recurrence cohort exhibited higher proportions of BPD (38.6% vs 15.6%, p<0.001), prior suicide attempts (46.2% vs 28.3%, p<0.001), and childhood trauma/abuse (58.6% vs 39.5%, p<0.001). The recurrence cohort also demonstrated greater acute behavioral instability, evidenced by more initial 24-h NSSI events (32.4% vs 11.8%, p<0.001) and elevated HAMD, HAMA, and CGI scores (all p<0.001). Additionally, baseline systemic inflammatory indices were significantly higher in the recurrence group, including NLR (2.85 vs 2.12, p<0.001), PLR (138.4 vs 115.6, p<0.001), and CRP (4.2 vs 2.5 mg/L, p<0.001). Metabolic panels and acute-phase pharmacological treatments were comparable between the two cohorts (all p>0.05).

Table 1 Demographic, Clinical, Laboratory, and Early in-Hospital Characteristics According to Observational Outcome

Cumulative Incidence of in-Hospital NSSI Recurrence and Competing Events

In the overall cohort (N=842), the cumulative incidence of NSSI recurrence was 8.5% (95% CI, 6.8–10.5%) at day 7 and 13.8% (95% CI, 11.6–16.2%) at day 14. The corresponding cumulative incidence of the competing event, predominantly survival discharge, was 18.2% (95% CI, 15.6–20.9%) and 48.5% (95% CI, 45.1–51.9%), respectively (Figure 1).

Figure 1 Cumulative incidence function curves for the overall study cohort during the 14-day prediction period. The solid line represents the cumulative incidence of in-hospital NSSI recurrence, and the dashed line represents the competing event, predominantly survival discharge without recurrence. Shaded areas indicate 95% confidence intervals.

For the exploratory categorical analysis, a maximally selected Gray statistic identified an NLR value of 2.45 as the data-derived cutpoint. Patients were consequently stratified into a high-NLR group (>2.45, n=398) and a low-NLR group (≤2.45, n=444). The cumulative incidence of NSSI recurrence was higher in the high-NLR group than in the low-NLR group at both day 7 (11.5% vs 5.8%) and day 14 (17.5% vs 10.5%; Gray’s test p<0.001) (Figure 2). The cumulative incidence of the competing event did not differ significantly between the two NLR strata during the 14-day prediction period (Gray’s test p=0.152).

Figure 2 Cumulative incidence of in-hospital NSSI recurrence during the 14-day prediction period according to baseline neutrophil-to-lymphocyte ratio (NLR). For this exploratory analysis, patients were categorized into a high-NLR group (>2.45) and a low-NLR group (≤2.45) using the data-derived cutpoint selected by the maximally selected Gray statistic. The cumulative incidence of recurrence was higher in the high-NLR group at both 7 and 14 days (Gray’s test, p<0.001). Shaded areas indicate 95% confidence intervals. The black dashed vertical lines indicate the prespecified 7-day and 14-day prediction horizons, measured from the 24-hour landmark.

Risk Factors for in-Hospital NSSI Recurrence and Non-Linear Analysis

Restricted cubic spline analysis demonstrated a significant non-linear association between baseline NLR and the risk of in-hospital NSSI recurrence (P for overall association <0.001; P for non-linearity =0.014). The estimated subdistribution hazard increased more steeply across the lower-to-middle NLR range and showed a more gradual increase at higher values (Figure 3A). In contrast, baseline HAMD score showed an approximately linear positive association with recurrence risk (P for overall association <0.001; P for non-linearity =0.385) (Figure 3B).

Figure 3 Restricted cubic spline (RCS) curves for the associations of baseline continuous variables with the risk of NSSI recurrence. (A) The non-linear association between continuous baseline NLR and the subdistribution hazard ratio (sHR) of NSSI recurrence (P for non-linearity =0.014). The solid line represents the estimated sHR, and the shaded area represents the 95% CI. The reference value was set at the median NLR of the overall cohort. (B) The linear dose-response relationship between baseline HAMD score and the sHR of NSSI recurrence (P for non-linearity = 0.385).

The primary multivariable Fine–Gray model included age, sex, primary BPD diagnosis, lifetime history of prior suicide attempt, lifetime history of childhood trauma or abuse, multiple methods of self-harm, NSSI during the first 24 hours, HAMD score, continuous NLR modeled using restricted cubic splines, and crisis intervention during the first 24 hours. NSSI during the first 24 hours (sHR, 2.12; 95% CI, 1.53–2.95; p<0.001), primary BPD diagnosis (sHR, 1.84; 95% CI, 1.32–2.57; p<0.001), childhood trauma or abuse (sHR, 1.62; 95% CI, 1.18–2.23; p=0.003), prior suicide attempt (sHR, 1.40; 95% CI, 1.03–1.90; p=0.031), and HAMD score (sHR per 1-point increase, 1.05; 95% CI, 1.02–1.08; p=0.001) were independently associated with subsequent recurrence. Continuous NLR remained associated with recurrence overall and showed a significant non-linear component (P for overall association <0.001; P for non-linearity =0.014) (Table 2).

Table 2 Univariate and Multivariate Fine-Gray Proportional Subdistribution Hazard Models Predicting in-Hospital NSSI Recurrence

In the sensitivity analysis excluding NSSI during the first 24 hours, the associations of BPD (sHR, 1.78; 95% CI, 1.28–2.49; p<0.001), childhood trauma or abuse (sHR, 1.57; 95% CI, 1.14–2.16; p=0.006), prior suicide attempt (sHR, 1.37; 95% CI, 1.01–1.86; p=0.043), and HAMD score (sHR per 1-point increase, 1.04; 95% CI, 1.01–1.07; p=0.005) were generally maintained. Continuous NLR also retained an overall association with recurrence (P<0.001), with evidence of non-linearity (P=0.018). Detailed results are presented in Supplementary Table S1. GVIF assessment identified substantial collinearity among psychiatric severity measures and among leukocyte-derived inflammatory indices. HAMA score, CGI-Severity score, WBC count, neutrophil count, lymphocyte count, PLR, CRP, and ESR had adjusted GVIF values ≥2.24 in the initial candidate model and were not entered simultaneously with HAMD score and NLR. All predictors retained in the final multivariable model had adjusted GVIF values <1.60 (Supplementary Table S2). Time-varying coefficient tests showed no evidence that the effect of any retained predictor varied significantly over follow-up (all p>0.05), and the joint test of all time-dependent coefficients was also non-significant (χ2=7.84, df=10, p=0.645), supporting the proportional subdistribution hazards assumption (Supplementary Table S3).

Construction and Internal Validation of the Competing Risk Nomogram

For clinical simplification, a secondary Fine–Gray model replaced the continuous NLR spline term with the exploratory dichotomous variable NLR >2.45. The resulting nomogram included primary BPD diagnosis, childhood trauma or abuse, prior suicide attempt, NSSI during the first 24 hours, HAMD score, and NLR >2.45 and estimated 7- and 14-day recurrence probabilities (Figure 4 and Supplementary Table S4). The apparent competing-risk C-index was 0.765 (95% CI, 0.728–0.802). Across 1000 bootstrap resamples, the mean optimism was 0.014, resulting in an optimism-corrected C-index of 0.751. The optimism-corrected calibration slopes were 0.94 at 7 days and 0.93 at 14 days, with calibration-in-the-large values of −0.01 and −0.02, respectively. Bias-corrected calibration curves showed close agreement between predicted and observed cumulative incidence at both time points (Figure 5A). Decision curve analysis demonstrated a positive net benefit over the treat-all and treat-none strategies across threshold probabilities of approximately 18–45% for the 7-day prediction and 14–45% for the 14-day prediction (Figure 5B).

Figure 4 Competing risk nomogram for predicting the 7-day and 14-day probability of in-hospital NSSI recurrence. To use the nomogram, a specific point value is assigned to each individual patient characteristic by drawing a vertical line upward to the “Points” axis. The sum of these points is located on the “Total Points” axis, and a vertical line is drawn downward to determine the patient’s estimated probability of NSSI recurrence at 7 and 14 days.

Figure 5 Internal validation and clinical utility of the competing-risk nomogram. (A) Optimism-corrected calibration curves for the predicted 7-day (red line) and 14-day (blue line) probabilities of in-hospital NSSI recurrence, obtained using 1000 bootstrap resamples. The x-axis represents the nomogram-predicted probability, and the y-axis represents the observed cumulative incidence estimated using the cumulative incidence function. The dashed 45-degree line represents perfect calibration. The bias-corrected curves showed close agreement between the predicted and observed cumulative incidence at both time points. (B) Decision curve analysis of the nomogram for the 7-day and 14-day prediction horizons. The solid red and blue lines represent the net benefit of the nomogram across different threshold probabilities. The nomogram provided greater net benefit than the treat-all and treat-none strategies across threshold ranges of approximately 18–45% for the 7-day prediction and 14–45% for the 14-day prediction.

Discussion

In this single-center retrospective cohort, we used a competing-risk framework to investigate subsequent in-hospital NSSI recurrence among psychiatric inpatients. The 14-day cumulative incidence of recurrence was 13.8%. The clinically prespecified multivariable Fine–Gray model identified NSSI during the first 24 hours, BPD, childhood trauma or abuse, prior suicide attempts, HAMD score, and NLR as predictors of subsequent recurrence. Continuous NLR showed a non-linear association with recurrence risk, whereas HAMD score showed an approximately linear association. The resulting nomogram demonstrated moderate discrimination, with an apparent C-index of 0.765 and an optimism-corrected C-index of 0.751. These findings support the potential value of combining clinical history, early behavioral information, symptom severity, and inflammatory markers, but they do not establish that the model is ready for routine clinical implementation.

The observed 14-day recurrence risk reflects the high clinical vulnerability of psychiatric inpatients with a recent history of NSSI. A large Chinese study reported an NSSI prevalence of 14.3% among adolescent psychiatric patients,18 whereas community-based research has reported a lifetime prevalence of approximately 19.3% among adolescents and young adults.19 These estimates are not directly comparable with the present cumulative incidence of recurrent in-hospital events, but they illustrate the substantial burden of NSSI across clinical and community settings. BPD and childhood trauma or abuse remained associated with recurrence after multivariable adjustment. This is consistent with evidence that childhood adversity may contribute to BPD-related emotional dysregulation and impulsivity.20 Neurobiological and psychological models further suggest that self-injury may function as a maladaptive response to hyperarousal and impaired stress regulation following early-life adversity.21 Nevertheless, the present observational findings demonstrate associations and should not be interpreted as establishing causal pathways.

The non-linear association between NLR and recurrence represents a potentially informative finding. Previous evidence has linked elevated NLR with suicidal behavior in patients with depression22 and with NSSI among adolescents with major depressive disorder.7 Peripheral inflammatory activity has been proposed to influence emotion regulation and behavioral control through immune–brain signaling and alterations in prefrontal–limbic circuits.23,24 However, the current data do not demonstrate that inflammation causes NSSI recurrence or that NLR constitutes a biological endophenotype. NLR is a nonspecific marker that may also reflect infection, physiological stress, psychological stress, medication exposure, or other unmeasured conditions. The continuous RCS analysis indicated that the association was non-linear, while the value of 2.45 was obtained from an exploratory maximally selected Gray statistic. This data-derived cutpoint should therefore be regarded as a cohort-specific aid to model simplification rather than a confirmed biological threshold.

NSSI during the first 24 hours was the strongest predictor of subsequent recurrence. This result is clinically expected because an early event identifies a patient who is already experiencing an active self-injury crisis and acute behavioral instability.25 Although early and subsequent events were temporally separated, this predictor provides limited lead time and primarily reflects persistence of an established crisis state rather than early warning before crisis onset. It should therefore not be presented as a novel etiological predictor. In the sensitivity analysis excluding this variable, the remaining associations of BPD, childhood trauma or abuse, prior suicide attempts, HAMD score, and NLR were generally maintained, and the C-index decreased from 0.765 to 0.741. Thus, the reduced model retained moderate predictive information among patients without directly observed early in-hospital self-injury. In this context, the non-linear NLR association may offer additional information because it is not itself a manifestation of the behavioral outcome and has previously been associated with self-injury and suicidal behavior.7,22

Crisis intervention was associated with recurrence in univariable analysis but not after multivariable adjustment. This pattern is compatible with confounding by indication: patients with greater behavioral instability and perceived risk are more likely to receive immediate crisis intervention.26,27 The unadjusted association should therefore not be interpreted as evidence that crisis intervention increases recurrence risk. More broadly, treatment-related associations in retrospective clinical records require caution because treatment assignment is determined by clinical need rather than random allocation.

The competing-risk framework was a methodological strength. Treating discharge without recurrence as a competing event avoided the upward bias that may occur when Kaplan–Meier methods censor patients who can no longer experience an in-hospital event.13 The Fine–Gray model therefore provided estimates that were more closely aligned with the clinical probability of recurrence before discharge.28 The analysis also incorporated non-linearity assessment, clinically prespecified predictors, GVIF-based collinearity evaluation, time-varying coefficient tests, multiple imputation, and bootstrap optimism correction. DCA suggested that the nomogram may provide greater net benefit than treat-all or treat-none strategies over selected threshold ranges.29 However, DCA evaluates potential decision utility under specified assumptions and does not demonstrate that model-guided care improves outcomes, reduces workload, or allocates observation resources safely. The observed discrimination was moderate, and the consequences of false-positive and false-negative classifications were not directly assessed. The nomogram should therefore be considered a research-stage adjunct to structured clinical assessment rather than a replacement for professional judgment.

Several limitations warrant consideration. First, this was a retrospective study from a single psychiatric center in China, and the cohort was young and predominantly female. Differences in age distribution, sex, diagnosis, admission criteria, staffing patterns, treatment pathways, and cultural context may limit generalizability to broader inpatient psychiatric populations. External validation in geographically and clinically diverse cohorts is required. Second, residual confounding remains possible because dynamic family circumstances, interpersonal conflicts, ward-level exposures, and the content and intensity of psychotherapeutic interventions could not be fully captured from electronic records. Third, NLR was measured only once within the first 24 hours. Because leukocyte counts may change with infection, stress, and medication exposure, this value represents an initial inflammatory snapshot rather than a longitudinal measure. Serial measurements are needed to determine whether NLR trajectories provide incremental predictive information. Fourth, the NLR cutpoint of 2.45 was data-derived and may overstate between-group separation or perform differently in other cohorts. External studies should reassess both the continuous functional form and the stability of this cutpoint. Fifth, the early NSSI predictor is conceptually close to the recurrence outcome and may enhance model performance by identifying an already active crisis. Although the sensitivity analysis supported the contribution of the remaining predictors, the primary model is better interpreted as short-term risk updating after initial observation than as prediction before any acute event. Finally, only internal validation was performed. Prospective multicenter validation, model updating where necessary, and implementation studies evaluating patient safety, staff workload, and resource use are required before clinical adoption.

Conclusion

In this single-center retrospective cohort, clinical history, depressive symptom burden, early in-hospital self-injury, and baseline NLR were associated with subsequent NSSI recurrence under a competing-risk framework. The association between NLR and recurrence was non-linear, but the exploratory cutpoint of 2.45 and the proposed biological interpretation require independent confirmation. NSSI during the first 24 hours was a strong but clinically expected predictor that primarily reflected an already active crisis and offered limited advance warning. The internally validated nomogram showed moderate discrimination and potential net benefit but should currently be regarded as a research-stage risk-support model rather than a ready-to-use clinical tool. External prospective validation in more diverse inpatient populations is required before it can inform patient-level interventions or resource allocation.

Data Sharing Statement

The experimental data used to support the findings of this study are available from the corresponding author upon request.

Ethics Approval

The study protocol was reviewed and approved by the Institutional Review Board of Hebei Provincial Mental Health Center (Approval No. 202012). The requirement for individual informed consent was waived by the Institutional Review Board because this was a retrospective, non-interventional analysis of existing de-identified medical records that posed minimal risk to participants. The waiver also applied to participants younger than 18 years; therefore, separate parental or legal guardian consent was not required for inclusion in this retrospective analysis. All data were handled confidentially, and the study was conducted in accordance with the Declaration of Helsinki.

Funding

The work was not funded by any funding.

Disclosure

The authors declared that they have no conflicts of interest regarding this work.

References

1. Apicella M, Pontillo M, Maglio G, et al. Non-suicidal self-injury in adolescents: a clinician’s guide to understanding the phenomenon, diagnostic challenges, and evidence-based treatments. Front Psychiatry. 2025;16:1605508. doi:10.3389/fpsyt.2025.1605508

2. Ye Z, Xiong F, Li W. A meta-analysis of co-occurrence of non-suicidal self-injury and suicide attempt: implications for clinical intervention and future diagnosis. Front Psychiatry. 2022;13:976217. doi:10.3389/fpsyt.2022.976217

3. Zhu W, Cui L, Zhang H, et al. Prospectively predicting 6-month risk for non-suicidal self-injury among adolescents after psychiatric hospitalization based on a predictive model. Front Psychiatry. 2024;15:1440808. doi:10.3389/fpsyt.2024.1440808

4. Liu PL, Zhang Y, Li J, Yang N, Du J, Dong QL. Non-suicidal self-injury in adolescent depression: a bibliometric study and visualization analysis. Acta Psychol. 2025;259:105306. doi:10.1016/j.actpsy.2025.105306

5. Liang S, Li D, Liu X, et al. Development and validation of a prediction nomogram for non-suicidal self-injury in female patients with mood disorder. Front Psychiatry. 2023;14:1130335. doi:10.3389/fpsyt.2023.1130335

6. Fasolato R, Raffagnato A, Miscioscia M, Gatta M. Non-suicidal self-injury in inpatient and outpatient adolescents: disentangling psychopathology and interactive family dynamics. Front Psychiatry. 2025;15:1483745. doi:10.3389/fpsyt.2024.1483745

7. Zhang W, Zhang Y, Xu T, Wang Y, Yu E, Yue J. Inflammatory indices are associated with non-suicidal self-injury in adolescents with major depressive disorder. BMC Psychiatry. 2025;26(1):25. doi:10.1186/s12888-025-07705-3

8. Han X, Yu X, Hu X, et al. C-reactive protein predicts aggression in adolescent depression: non-suicidal self-injury moderation and immunomodulatory shift. Psychoneuroendocrinology. 2026;186:107756. doi:10.1016/j.psyneuen.2026.107756

9. Baydili İ, Tasci B, Tasci G. Artificial intelligence in psychiatry: a review of biological and behavioral data analyses. Diagnostics. 2025;15(4):434. doi:10.3390/diagnostics15040434

10. Mürner-Lavanchy I, Koenig J, Reichl C, Josi J, Cavelti M, Kaess M. The quest for a biological phenotype of adolescent non-suicidal self-injury: a machine-learning approach. Transl Psychiatry. 2024;14(1):56. doi:10.1038/s41398-024-02776-4

11. Hooley JM, Fox KR, Boccagno C. Nonsuicidal self-injury: diagnostic challenges and current perspectives. Neuropsychiatr Dis Treat. 2020;16:101–13. doi:10.2147/NDT.S198806

12. Posner K, Brown GK, Stanley B, et al. The Columbia-suicide severity rating scale: initial validity and internal consistency findings from three multisite studies with adolescents and adults. Am J Psychiatry. 2011;168(12):1266–1277. doi:10.1176/appi.ajp.2011.10111704

13. Austin PC, Lee DS, Fine JP. Introduction to the analysis of survival data in the presence of competing risks. Circulation. 2016;133(6):601–609. doi:10.1161/CIRCULATIONAHA.115.017719

14. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40(5):373–383. doi:10.1016/0021-9681(87)90171-8

15. Busner J, Targum SD. The clinical global impressions scale: applying a research tool in clinical practice. Psychiatry. 2007;4(7):28–37.

16. Hamilton M. A rating scale for depression. J Neurol Neurosurg Psychiatry. 1960;23(1):56–62. doi:10.1136/jnnp.23.1.56

17. Hamilton M. The assessment of anxiety states by rating. Br J Med Psychol. 1959;32(1):50–55. doi:10.1111/j.2044-8341.1959.tb00467.x

18. Zhang P, Ouyang L, Liang M, et al. A cross-sectional epidemiological study of non-suicidal self-injury prevalence in Chinese psychiatric patients. Nat Mental Health. 2023;1(4):266–272. doi:10.1038/s44220-023-00050-y

19. Voss C, Hoyer J, Venz J, Pieper L, Beesdo-Baum K. Non-suicidal self-injury and its co-occurrence with suicidal behavior: an epidemiological-study among adolescents and young adults. Acta Psychiatr Scand. 2020;142(6):496–508. doi:10.1111/acps.13237

20. Dodd DR, Crosby RD, Cao L, Gordon KH, Wonderlich SA. Borderline personality disorder symptoms as mediational mechanisms linking childhood trauma and nonsuicidal self-injury among women with bulimia nervosa. Int J Eat Disord. 2022;55(3):372–381. doi:10.1002/eat.23669

21. Reitz S, Kluetsch R, Niedtfeld I, et al. Incision and stress regulation in borderline personality disorder: neurobiological mechanisms of self-injurious behaviour. Br J Psychiatry. 2015;207(2):165–172. doi:10.1192/bjp.bp.114.153379

22. Velasco A, Lengvenyte A, Rodriguez-Revuelta J, et al. Neutrophil-to-lymphocyte ratio, platelet-to-lymphocyte ratio, and monocyte-to-lymphocyte ratio in depressed patients with suicidal behavior: a systematic review. Eur Psychiatry. 2023;67:1–25. doi:10.1192/j.eurpsy.2023.18

23. Petruso F, Giff AE, Milano BA, De Rossi MM, Saccaro LF. Inflammation and emotion regulation: a narrative review of evidence and mechanisms in emotion dysregulation disorders. Neuronal Signal. 2023;7(4):NS20220077. doi:10.1042/NS20220077

24. Bauer ME, Teixeira AL. Neuroinflammation in Mood Disorders: role of Regulatory Immune Cells. Neuroimmunomodulation. 2021;28(3):99–107. doi:10.1159/000515594

25. McIvor L, Payne-Gill J, Beck A. Associations between violence, self-harm and acute psychiatric service use: implications for inpatient care. J Psychiatr Ment Health Nurs. 2023;30(3):451–460. doi:10.1111/jpm.12872

26. Lee J, Chang SM. Confounding by Indication in Studies of Selective Serotonin Reuptake Inhibitors. Psychiatry Invest. 2022;19(11):873–883. doi:10.30773/pi.2022.0209

27. Papola D, Ostuzzi G, Gastaldon C, et al. Antipsychotic use and risk of life-threatening medical events: umbrella review of observational studies. Acta Psychiatr Scand. 2019;140(3):227–243. doi:10.1111/acps.13066

28. Mirian C, Jensen LR, Juratli TA, et al. The importance of considering competing risks in recurrence analysis of intracranial meningioma. J Neurooncol. 2024;166(3):503–511. doi:10.1007/s11060-024-04572-y

29. Berlow YA, Zandvakili A, Philip NS. The clinical utility of imaging-defined biotypes of depression and transcranial magnetic stimulation: a decision curve analysis. Brain Stimul. 2020;13(4):1069–1070. doi:10.1016/j.brs.2020.04.016