Detecting and distinguishing indicators of risk for suicide using clinical records

Study sample description

Summary statistics for demographic variables are shown in Table 1. Overall, cases were older (mean age for cases was 51.4 years and 48.1 years for controls) and more likely to be male (cases 77.5% and controls 46.2%). Cases were less likely to have commercial insurance (cases 62.8% and controls 72.5%), more likely to live in areas with lower education levels (cases 39.1% and controls 36.9%), and had lower neighborhood household incomes (case median = $65,567 and control median = $67,694). Table 1 also shows that case/control summary statistics were consistent between discovery and validation samples.

Table 1 Demographic characteristics of study subjects by suicide case/control status.

Single healthcare indicator associations with suicide death and heterogeneity by MH status

The single indicator association results are included in Supplementary Table 1. Overall, of the 202 indicators tested in the discovery sample, 170 were significantly associated with suicide death (FDR < 0.05), and of these, 146 (86%) were also significant in the validation sample, including 78 diagnoses (77 increased/1 decreased risk), 45 procedures (44/1), and 21 encounter types (21/0). The top ten most significant, validated associations for each indicator type are summarized in Table 2.

Table 2 Univariate odds ratios for death by suicide adjusted by age and sex.

All indicators were evaluated for differential suicide risk by MH status (Supplementary Table 2). Of the 202 individual healthcare indicators, 44 had significant interactions (FDR < 0.05) with MH status in the discovery set, with 7 (16%) also significant in the validation (p < 0.05), with MH-stratified results displayed in Table 3. Of note, malignant neoplasms were associated with >twofold higher risk among those without MH diagnoses. Given these differences, subsequent multi-indicator analyses were stratified by MH status.

Table 3 Mental health stratified odds ratios for validated interactions between utilization features and mental health status.

Multi-indicator associations with suicide by MH status

The results from the discovery/validation penalized conditional logistic regression models stratified by MH status are displayed in Supplementary Table 3. The MH model retained 87 indicators (46 diagnoses, 30 procedures, and 11 encounters), and in the validation model, 49 indicators (24 diagnoses, 19 procedures, and 6 encounters) were retained. Increased/decreased risk of suicide was significantly associated with 7/8 diagnoses, 9/2 procedures, and 2/0 encounters. The non-MH discovery model retained 98 indicators (56 diagnoses, 29 procedures, and 13 encounters), and in the corresponding validation model, 51 indicators (25 diagnoses, 8 procedures, and 18 encounters) were retained. Increased/decreased risk was significantly associated with 7/4 diagnoses, 5/4 procedures, and 2/2 encounters.

The full MH and non-MH samples were used to estimate 95% CIs and significance for indicators retained in the respective validation models, and the results for those indicators that remained statistically significant are displayed in Fig. 1. For diagnoses, several individual MH conditions increased risk, including Other Psychoses (OR = 4.00, 95% CI 3.50–4.58) and Personality Disorders and Other Nonpsychotic Mental Disorders (OR = 3.58, 95% CI 3.00–4.29). In comparison, cancer diagnoses had a unique increased risk among non-MH individuals, including Malignant Neoplasms of Other and Unspecified Sites (OR = 4.29, 95% CI 2.95–6.23) and Malignant Neoplasm of Respiratory and Intrathoracic Organs (OR = 2.90, 95% CI 1.66–5.06). In contrast, Benign Neoplasms had protective effects for both non-MH (OR = 0.65, 95% CI 0.50–0.85) and MH (OR = 0.68, 95% CI 0.55–0.84) individuals. The services related to Reproduction and Development were protective in both models, although the effect was more extreme in the non-MH (OR = 0.37, 95% CI 0.24–0.58) relative to the MH (OR = 0.60, 95% CI 0.45–0.81) model.

Fig. 1: Forest plots of odds ratios from multi-indicator models of suicide death among those A) with and B) without a mental health diagnosis.
figure 1

Results in both panels are taken from conditional logistic regression models fit to the full sample of those individual with and without mental health diagnoses. For each model, indicators were selected from those that were retained in both the discovery and validation penalized regression models, and only those that were statistically significant (p < 0.05) in the full sample were included in these figures.

For procedures, increased risk among MH individuals was distinguished by procedures related to the Respiratory System (OR = 1.95, 95% CI 1.50–2.52) and Psychiatry (OR = 1.24, 95% CI 1.08–1.43). In comparison, Radiology procedures (OR = 1.54, 95% CI 1.14–2.08) were significant for the non-MH model. The MH sample had a unique protective effect of Nursing Facility Services (OR = 0.54, 95% CI 0.36–0.81). Non-MH individuals had unique protective effects of Immunization Administration for Vaccines/Toxoids (OR = 0.82, 95% CI 0.68–0.98), Preventive Medicine Services (OR = 0.74, 95% CI 0.62–0.88), and Ophthalmology (OR = 0.76, 95% CI 0.64–0.90).

For encounters, increased risk among MH individuals was distinguished by Non-Face-to-Face Non-physician Services (OR = 1.53, 95% CI 1.13–2.07), Acute Inpatient (OR = 1.47, 95% CI 1.21–1.79), and Ambulatory – Rehab (OR = 1.53, 95% CI 1.16–2.02). For non-MH individuals, Other Non-overnight – Home Health (OR = 1.53, 95% CI 1.16–2.00) and Radiology Only – Outpatient Clinic (OR = 1.54, 95% CI 1.14–2.08) were associated with increased risk. MH individuals had no encounter types with protective effects, and non-MH individuals had protective effects for Emergency – Hospital Ambulatory (OR = 0.53, 95% CI 0.40–0.70) and Email – Other Non-hospital (OR = 0.68, 95% CI 0.52–0.88).

Latent class analysis discovery and validation of distinct patient suicide risk sub-groups

The top indicators that differentiate the LCA identified sub-groups for both MH and non-MH samples are displayed in Fig. 2. For both the MH and non-MH samples, the number of latent subgroups was identified as five, based on lowest value past the inflection point in the BIC curve (Supplementary Fig. 2) that also identified a low-risk group. For the MH sample, the groups are labelled in order of decreasing case percentage: Group 1 (10.0% of the sample, 13.8% cases), Group 2 (16.1% of the sample, 11.0% cases), Group 3 (26.0% of the sample, 4.1% cases), Group 4 (24.1% of the sample, 3.6% cases), and Group 5 (23.8% of the sample, 1.9% cases). Based on these values, Groups 1 and 2 were identified as high-risk groups for suicide. Specifically, Group 1 had higher proportions of many diagnosis sub-chapters, procedure types, and encounter types, signifying a high utilization group with multiple healthcare concerns. Group 2 had a similar but less extreme profile, and these individuals were also younger (39 years) and more likely to be female. Group 5 had the lowest case prevalence. This group was one of the youngest on average (44 years old), had the lowest proportion of males, and the highest proportion of routine/preventive health visits. Confirmatory LCA of the five-group solution in the validation sample yielded groups with similar case percentages and healthcare indicator profiles (Fig. 3).

Fig. 2: Latent class analysis sub-group identification based on health care indicators associated with suicide death in individuals A) with and B) without a mental health diagnosis.
figure 2

For each mental health stratum, latent class analysis (LCA) was performed based on health care indicators identified in the respective penalized regression models in the discovery sample. LCA was performed in the discovery sample (D), followed by confirmatory LCA in the validation sample (V). The frequency of cases is displayed overall and for each LCA sub-group for both D and V samples. Within each stratum, the LCA sub-group-specific frequency are displayed for those health care indicators where at least 9 of the 10 ratios of the pairwise group frequencies were >1.5 (ie. sub-group distinguishing indicators).

For non-MH individuals, he resulting groups and their corresponding case proportions were: Group 1 (14.1% of the sample, 2.7% cases), Group 2 (21.8% of the sample, 1.7% cases), Group 3 (25.6% of the sample, 1.5% cases), Group 4 (25.3% of the sample, 1.3% cases), and Group 5 (13.3% of the sample, 0.3% cases). The noticeably high-risk Group 1 was the second oldest (59 years old) and contained much higher percentages of most healthcare indicators with multiple concerns, similar to Groups 1 and 2 from the MH sample. Further, while Groups 3 and 4 where at intermediate risk and had similar but less extreme healthcare profiles, there was an additional disengaged, high risk group (Group 2) unique to the non-MH sample. The lowest-risk Group 5 was the youngest (39 years), contained a low proportion of males (5.6%), and displayed higher proportions of routine/preventive visits – similar to Group 5 from the MH model. Confirmatory LCA of the five-group solution resulted in both similar case proportions and healthcare indicator profiles.

Source link

Leave A Reply

Your email address will not be published.