Predictive models for active suicidal ideation in cognitive decline: identifying risk factors

Abstract

Background

Suicide rates among older adults with cognitive decline represent a critical public health concern. Despite the association between cognitive decline and suicidality, predictive models for active suicidal ideation (ASI) in this population remain underexplored.

Methods

A retrospective study of 1,889 patients with cognitive decline was conducted using electronic health records. Sociodemographic, cognitive, clinical, psychiatric, and functional variables were analyzed. Univariate logistic regression identified correlates of ASI, followed by multivariate predictive modeling using Logistic Regression (LR) and XGBoost. Recursive Feature Elimination (RFE) identified key predictors, and SHAP values provided model interpretability.

Results

Depressive symptoms, Mini-Mental-State-Exam score, duration of cognitive decline, past suicide attempts, antidementia medication use, and living arrangement emerged as key predictors. Both LR and XGBoost demonstrated robust performance (ROC AUC: 0.81–0.70; PR AUC: 0.55).

Conclusion

Multivariate predictive models provide improved risk stratification for ASI, highlighting the need for targeted interventions among individuals with cognitive decline.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-025-07768-2.

Background

Suicide remains a significant global health concern across all age groups, with rates increasing notably among older adults [1]. In Slovenia, the suicide rate in 2022 was 19.06 per 100,000 individuals, with a marked increase among older adults, ranging from 25 to 39 per 100,000 for both sexes as age progresses beyond 50 [2]. As the older population continues to grow and social risk factors become more prevalent, suicide behavior among the elderly is expected to become an even greater public health concern.

Suicide is often conceptualized as a continuum, beginning with thoughts about death and progressing to planning, intent, attempts, and ultimately, completed suicides [3, 4]. Building on this view, ideation-to-action theories have expanded this understanding by distinguishing between the factors that give rise to suicidal thoughts and those that enable their enactment. These models propose that pain, hopelessness, and perceived burdensomeness contribute to suicidal ideation, whereas the capacity to act on such thoughts (reduced fear, planning ability access to means) determines risk for suicidal behavior [5]. This continuum highlights the importance of early identification and intervention, particularly for individuals with active suicidal ideation (ASI) that involves planning or intent, to prevent escalation to more severe outcomes [6]. ASI represents a clinically urgent and high-risk stage of suicidal thinking, strongly associated with suicide attempts [7]and requiring prompt detection and management. However, current suicide risk prediction models often rely on straightforward clinical indicators, such as previous suicide attempts, mental health diagnoses, or demographic factors, which can be too simplistic [8–10]. This oversimplification may result in inadequate risk assessments and missed opportunities for timely intervention.

Researchers have identified several risk factors for suicidal ideation and behavior in older adults [11, 12]. However, “Old age” may be more of a social construct than a meaningful biological threshold. A more nuanced understanding of aging incorporates cognitive or functional decline, which better captures the complexities of the aging process [13]. Cognitive decline, in particular, has been recognized as a significant risk factor for suicidal ideation and behavior, with many suicide attempts occurring within the first year following a diagnosis [14, 15]. Although many studies have examined risk factors and correlates of suicide in older adults [3, 4, 16], few have employed predictive modeling approaches [17–19]. Moreover, most existing suicide risk prediction models have been developed primarily in non-geriatric samples, with limited applicability to older adults [20–25].

To address these gaps, this study focuses on individuals with cognitive decline by examining the relationships between clinical, cognitive, and functional factors in this population. In contrast to most previous work that considers suicidal ideation more broadly, this study specifically targets active suicidal ideation, a more severe and clinically meaningful form of suicidal thinking. By integrating these factors within a predictive framework, the study aims to identify key risk markers and improve the precision of suicide risk assessment. Ultimately, it seeks to support the development of models that enable accurate risk stratification and facilitate timely, targeted interventions for those at greatest risk.

Methods

The study was approved by the Medical Ethics Committee of the University Psychiatric Clinic Ljubljana and conducted in accordance with the ethical standards of the 1964 Declaration of Helsinki and its later amendments. Participants’ data were anonymized accordingly.

This retrospective study included 1,889 patients (1,106 females and 783 males) diagnosed with cognitive decline admitted to the University Psychiatric Clinic Ljubljana between January 2019 and April 2024. Diagnoses were classified according to the International Classification of Diseases, 10th Revision (ICD-10: F00, F01, F02, F03, F06.7, G30). We adopted a clinical, transdiagnostic perspective, treating cognitive decline as a shared syndrome with overlapping cognitive and affective features across etiologies. To stratify suicide risk based on suicidal ideation, we categorized patients according to their responses on the Columbia-Suicide Severity Rating Scale (C-SSRS) quick screen. ASI was considered indicative of a higher suicide risk and was defined by a “yes” response to any of the following C-SSRS questions: “Have you been thinking about how you might kill yourself?“, “Have you had these thoughts and had some intention of acting on them?“, or “Have you started working out the details of how to kill yourself? Do you intend to carry out this plan?“. Patients with ASI were grouped into the ASI group, while those with no or only passive suicidal ideation were assigned to the no or passive suicidal ideation group (NPSI).

Data collection

Data were extracted from electronic health records, including sociodemographic, cognitive, clinical, psychiatric, and functional variables. Sociodemographic variables recorded were sex, age, living arrangement, categorized as living in a shared household, living alone, residing in a nursing home, or another form of organized living, as well as the highest level of formal education attained. Cognitive variables included the type of cognitive decline, the duration of cognitive decline from symptom onset and diagnosis, and scores from the Mini-Mental State Examination (MMSE) and the Clock Drawing Test (CDT). Clinical variables included the presence of depressive symptoms at admission, use of antidementia medication (ADMs), such as acetylcholinesterase inhibitors and N-Methyl-D-Aspartate receptor antagonists (NMDA), somatic comorbidities (total number of physical health conditions diagnosed), reported pain, and body mass index (BMI). Psychiatric variables included a history of previous psychiatric disorders, categorized into affective disorders, psychotic, and anxiety spectrum disorders, as well as a history of substance abuse. For these binary variables, absence of documentation in the medical record was coded as ‘no’. Additionally, data were collected for all subjects regarding any history of previous suicide attempts. Functional status was assessed by physiotherapists using a series of standardized tests, including the Sit-to-Stand Test (STS), Functional Reach Test (FRT), grip strength, 4-meter and 10-meter walk tests (4mWT and 10mWT, respectively), 6-minute walk test (6MWT), Timed Up and Go Test (TUG), and the de Morton Mobility Index 100-point scale (DEMMI).

Statistical analysis and machine learning methods

First, univariate logistic regression was performed to identify correlates of ASI. Additionally, we report subgroup univariate analyses stratified by diagnosis (Alzheimer’s/MCI N = 1,216; vascular dementia N = 378; other dementias N = 302) to examine potential diagnostic differences; results are provided in the Supplement (Table S3). Secondly, the predictive power of single variables identified as significant in the univariate analysis was assessed using a logistic regression model. Before proceeding with multivariate analysis, collinearity tests were conducted for all variables. The duration of cognitive decline from symptom onset or diagnosis, as well as 4mWT, and 10mWT were identified as collinear. Given the importance of the duration variable, we implemented a combined “duration” variable, where either the duration since cognitive decline onset or, when unavailable, the duration since diagnosis was taken. The following section outlines the machine learning pipeline designed to extract complex patterns from the dataset. Additionally, the pipeline is visualized in Supplementary Information (Figure S2).

Due to the relatively large proportion of missing values in the dataset (see Table 1), a k-Nearest Neighbours (k-NN) imputation method was applied, with k = 5 to address the missing values effectively. The dataset was subsequently divided into two subsets: 80% for training and 20% for testing. Two machine learning models were employed: a multivariate logistic regression (LR) and an extreme gradient boosting model (XGBoost) to construct predictive models for suicide risk. We selected LR as an interpretable clinical baseline capable of capturing linear interactions between variables and XGBoost as a tree-boosting method capable of modeling nonlinear effects and feature interactions. We did not expand to additional algorithms (Random Forest, SVM) because, on structured tabular data, gradient-boosted trees typically perform at least comparably to other common baselines while offering interpretability via SHAP [24, 26]. Due to the presence of class imbalance within the dataset, we applied class weights to the LR model, proportionate to the imbalance between classes. For the XGBoost model, class imbalance was addressed using hyperparameter tuning optimising on ROC AUC and PR AUC using the full set of features. To enhance the robustness, hyperparameters were tuned using repeated stratified 5-fold cross-validation (10 repeats; 50 resamples in total), where each model was trained on 80% and validated on 20% of the training data. Performance metrics were averaged across all resamples to select the final configuration (max_depth: 2, min_child_weight: 9, n_estimators: 20, scale_pos_weight: 1, and subsample: 0.9). Next, Recursive Feature Elimination (RFE) was employed to determine the most predictive features. By varying the number of selected features, we identified that after 6 features, additional variables contributed minimal improvements to the model’s performance (see Fig. 1). To ensure robust feature selection, the RFE process was repeated across a hundred different partitions of the training data, selecting the features most frequently identified as informative (Fig. 1). Due to the plateauing performance, we decided to retain a simpler model with fewer features to improve interpretability and practicality. Including additional predictors would have inherently increased the proportion of missing data and thereby complicating the analysis.

Table 1.

Sociodemographic and clinical characteristics

Sociodemographic factors		Total n	Missing (%)	n_ASI (%)	n_NPSI (%)	OR (95% CI)	P-value
Sex		1896	0%				0.14
	F			129 (7%)	984 (52%)	1.26 (0.93, 1.70)
	M			74 (4%)	709 (37%)	0.8. (0.59, 1.08)
Age		1896	0%
				82 (77, 87.5); (11%)	83 (77, 87); (89%)	1.05 (0.91, 1.22)	0.50
Living arrangement		1893	0.2%
	Shared			95 (5%)	859 (45%)	0.98 (0.69, 1.40)	0.93
	Nursing home			44 (2%)	371 (20%)	0.65 (0.20, 2.13)	0.48
	Alone			61 (3%)	422 (22%)	1.29 (0.94, 1.78)	0.11
	Other			3(0.2%)	38 (2%)	0.65 (0.20, 2.13)	0.50
Education		1103	41.8%
				2 (1, 3); (11%)	2 (2, 3); 89%	0.91 (0.75, 1.10)	0.33
Cognitive status
Diagnosis		1896	0%
	MCI			30 (2%)	135 (7%)	2.00 (1.31, 3.06)	0.001
	AD			106 (6%)	945 (50%)	0.86 (0.65, 1.16)	0.32
	VD			38 (2%)	340 (18%)	0.92 (0.63, 1.33)	0.65
	UD			26 (1%)	234 (12%)	0.92 (0.59, 1.41)	0.69
	Other			3 (0.2%)	38 (2%)	0.63 (0.19, 2.07)	0.45
Duration Dx [yr]		1320	30.4%	0.00 (0, 1); (12%)	0.08 (0, 2); (87%)	0.75 (0.61, 0.94)	0.01
Duration SxS [yr]		1081	43%	0.5 (0, 2); (12%)	2 (0.16, 4.5); (88%)	0.59 (0.45, 0.78)	0.001
MMSE score		1557	17.9%	18 (14, 22); (12%)	15 (8, 20); (88%)	1.57 (1.34, 1,85)	< 0.001
CDT score		569	70%	1,5 (1, 3); (12%)	1 (0, 2); (88%)	1.50 (1.18, 1.88)	< 0.001
Clinical factors
Depressive symptoms		1889	0.4%
	Present			123 (7%)	236 (12%)	9.45 (6.90, 12.92)	< 0.001
	Absent			80 (4%)	1450 (77%)
ADMs		1876	1.1%
	Yes			59 (3%)	725 (39%)	0.55 (0.40, 0.75)	0.002
	No			141 (8%)	951 (51%)
Somatic comorbidities		1890	0.3%	4 (3, 6); (11%)	4 (2,5); (89%)	1.16 (1.01, 1.34)	0.03
Pain		1888	0.4%
	Present			68 (4%)	398 (21%)	1.63 (1.19, 2.23)	0.002
	Absent			135 (7%)	1287 (68%)
BMI		1298	31.5%	25.8 (22.5, 29.4); (11%)	25 (22.2, 28.4); (89%)	1.15 (0.98, 1.37)	0.08
Psychiatric history		1896	0%
Past SA	Yes			39 (2%)	50 (3%)	7.81 (4.99, 12.24)	< 0.001
	No			164 (9%)	1643 (87%)
Affective disorder	Yes			61 (3%)	274 (15%)	2.22 (1.60, 3.08)	< 0.001
	No			142 (7%)	1419 (75%)
Psychotic disorder	Yes			4 (0.2%)	77 (4%)	0.42 (0.14, 1.17)	0.096
	No			199 (10%)	1616 (85%)
Anxiety disorder	Yes			20 (1%)	135 (7%)	1.26 (0.77, 2.06)	0.36
	No			183 (10%)	1558 (82%)
Substance abuse	Yes			35 (2%)	193 (10%)	1.62 (1.10, 2.40)	0.02
	No			168 (9%)	1500 (79%)
Functional status
DEMMI [x/100]		1013	46.6%	67 (44, 74);(13%)	57 (39, 74); (87%)	1.33 (1.10, 1.61)	0.004
Grip strength [kg]		964	49.2%	17 (10, 22.5); (13%)	15 (9, 22); (87%)	1.18 (0.98, 1.41)	0.08
4mWT [s]		955	49.6%	6.79 (4.84, 12.45); (13%)	7.64 (5.32, 12.64); (87%)	1.10 (0.95, 1.28)	0.18
6MWT [m]		947	50.1%	205 +/- 126.57 (13%)	175 +/- 125.16 (87%)	1.12 (0.93, 1.34)	0.22
TUG [s]		904	52.3%	19.80 (12.6, 40.70); (14%)	21.89 (14.24, 36.34); (86%)	1.10 (0.93, 1.32)	0.25
10mWT [s]		959	49.7%	15.47 (10.84, 26.80); (14%)	17.59 (12.17, 28); (86%)	0.95 (0.78, 1.15)	0.61
FRT [cm]		954	49.7%	-30 (-38, -19); (14%)	-30 (-40, -18); (86%)	0.99 (0.83, 1.20)	0.94
STS		941	50.4%	5.0 (0, 9); (13%)	4.0 (0, 8); (87%)	1.25 (1.04, 1.50)	0.02
Walking ability		1108	41.6%
	Unaided walking			87 (8%)	577 (52%)	0.98 (0.69, 1.40)	0.93
	Rollator/Walker			40 (4%)	270 (24%)	0.97 (0.66, 1.43)	0.87
	Assistance			19 (2%)	115 (10%)	1.10 (0.66, 1.85)	0.71

Sociodemographic and clinical characteristics (ASI vs. NPSI). Data are reported as median (IQR; Q1–Q3); (n%)for continuous variables and n (%) for categorical variables, unless otherwise indicated

Abbreviations: ASI, active suicidal ideation; NPSI, no or passive suicidal ideation; MCI, mild cognitive impairment; AD, Alzheimer’s disease; VD, vascular dementia; UD, unspecified dementia; Duration Dx, duration since diagnosis; Duration SxS, duration since onset of cognitive-decline; MMSE score, Mini Mental State Exam; CDT, Clock Drawing Test; ADMs, anti-dementia medication; SA, suicide attempt; DEMMI, de Morton Mobility Index; 4MWT, 4 m Walk Test; 6MWT, 6 min Walk Test; FRT, Functional Reach Test; STS, Sit-to-Stand Test. Education (completed): 0 – no formal education, 1 – secondary education (8 year), 2 – vocational or technical training, 3 – high school or equivalent diploma, 4 – bachelor’s degree, 5 – master’s degree, 5 – PhD or equivalent advanced qualification

Fig. 1.

Recursive feature elimination. The figure on the left shows the relationship between model performance metrics (ROC AUC, top, and PR AUC, bottom) and the number of features used. After 6 features, adding more contributed little to improving model’s performance. The figure on the right shows the relative frequency of features selected during the feature selection process. Abbreviations: Past SA, past suicide attempt(s); MMSE, Mini-Mental-State-Exam; ADMs, antidementia medication; FRT [cm], Functional Reach Test; 10mWT [s], 10 m Walk Test

With the top six features selected, we trained both the XGBoost and LR models without performing any data imputation, as the selected features, leaving us with N = 1184 complete cases (143 ASI individuals). The training and test set thus comprised cases from the original training and test sets that had values for the selected features. The training set included 947 participants (110 ASI; 11.6%), and the test set 237 participants (33 ASI; 13.9%). We then trained both models using the previously optimized hyperparameters and evaluated their performance, reporting the ROC AUC and PR AUC values for both models (for the test set, see Results section; for the train set, see SI Table S2). In addition, SHAP (SHapley Additive exPlanations) values were computed to provide insights into the contributions of individual variables to the model predictions, offering a clear interpretation of how each feature influenced the output of XGBoost.

Results

Correlates of active suicidal ideation and their discriminatory ability

In the univariate analysis, several factors were found to be significantly associated with ASI, as represented in Table 1, and Figure S1 in the Suppl. Sociodemographic factors, such as age, living arrangement, and the highest level of formal education attained, did not show a statistically significant association with the risk of ASI. Cognitive variables showed significant associations. Higher scores on the MMSE and CDT were linked to an increased risk of ASI (OR: 1.57, 95% CI: 1.34–1.85, p < 0.001; OR: 1.50, 95% CI: 1.18–1.88, p < 0.001, respectively). MCI was similarly associated with a higher risk (OR: 2.00, 95% CI: 1.31–3.06, p = 0.001). Conversely, the duration since cognitive decline onset was negatively associated with ASI (OR: 0.59, 95% CI: 0.45–0.78, p = 0.001), as was the duration since diagnosis (OR: 0.75, 95% CI: 0.61–0.94, p = 0.01). Within the clinical domain, depressive symptoms at admission showed the strongest correlation, with an OR of 9.45(95% CI: 6.90–12.92, p < 0.001). Pain was positively associated with ASI (OR: 1.63, 95% CI: 1.19–2.23, p = 0.002), and there was a modest positive correlation with the somatic comorbidities (OR: 1.16, 95% CI: 1.01–1.34, p = 0.03). In contrast, the use of ADMs was associated with a moderately reduced risk (OR: 0.55, 95% CI: 0.40–0.75, p = 0.002). Previous mood disorders were significantly linked to a higher risk of ASI (OR: 2.22, 95% CI: 1.60–3.08, p < 0.001), as was a history of substance abuse (OR: 1.62, 95% CI: 1.10–2.40, p = 0.02). A history of previous suicide attempts was also a significant predictor of ASI (OR: 7.81, 95% CI: 4.99–12.24, p < 0.001). In terms of functional status, higher DEMMI scores were significantly associated with an increased risk of ASI (OR: 1.33, 95% CI: 1.10–1.61, p = 0.004). Similarly, a better performance on the STS test was also associated with ASI (OR: 1.25, 95% CI: 1.04–1.50, p = 0.02).

Next, the discriminatory power of significant predictors for ASI was evaluated using ROC AUC and PR AUC values. Among the variables, depressive symptoms at admission (ROC AUC = 0.70, PR AUC = 0.46) and the duration since symptom onset (ROC AUC = 0.67, PR AUC = 0.36) emerged as the strongest predictors. However, most predictors exhibited limited discriminatory power, with ROC AUC values from 0.4 to 0.6 (Table S1 in Suppl.).

Model performance and key predictors of active suicidal ideation

To explore more complex relationships among these correlates, predictive models were developed using LR and XGBoost. First, RFE identified the most predictive features, including depressive symptoms at admission, past suicide attempts, duration of cognitive decline, as well as MMSE score, use of ADMs, and living arrangements (Fig. 1). On these predictive features both LR and XGBoost demonstrated good discriminative ability for assessing the risk of ASI with a ROC AUC of 0.79 and 0.71, respectively. They also showed a good ability to identify the positive class in an imbalanced set (PR AUC 0.55 for both, with a baseline of 0.16) (Table 2; Fig. 2). Regarding precision, sensitivity, and F1 score, LR showed better performance than XGBoost (Table 2). Furthermore, SHAP analysis provided further insight into how these features contributed to the model’s predictions (Fig. 3). The presence of depressive symptoms at admission had the most substantial positive impact on predicting ASI, whereas their absence corresponded to a negative impact on the model’s output. Shorter durations of cognitive decline showed a positive influence on the model’s output, whereas longer durations had a stronger negative impact. MMSE scores exhibited a complex relationship with the model predictions; higher MMSE scores generally resulted in positive SHAP values, with medium-to-high scores presenting the highest positive SHAP value, suggesting an increased risk, while lower scores corresponded with negative SHAP values, indicating a lower risk. The use of ADMs was associated with a slight decrease in risk, as reflected by negative SHAP values for individuals taking these medications. For patients with a history of suicide attempts, SHAP values were consistently positive, underscoring the increased risk associated with this factor. Living arrangements also influenced the model’s outcome. Residing in a nursing home was associated with a positive contribution while living alone had values centered around zero.

Table 2.

Performance of LR and XGBoost models on the test set

	ROC AUC	PR AUC	Precision	Sensitivity	F1	NPV
LR	0.79	0.55	0.4	0.70	0.51	0.94
XGBoost	0.81	0.55	0.55	0.48	0.52	0.92

Abbreviations: LR, Logistic regression model; XGBoost, Extreme Gradient Boosting model; NPV, Negative predictive value

Fig. 2.

ROC (left) and PR curves (right) for XGBoost and LR models. Both models achieved good discriminatory ability, XGBoost achieved a ROC AUC of 0.81, and LR a ROC AUC of 0.79. Both models achieved PR AUC scores of 0.55, significantly above the baseline of 0.16, reflecting effective handling of the minority class

Fig. 3.

SHAP values illustrate the contribution of individual features to the model’s prediction of active suicidal ideation (ASI). Each dot represents a patient, with its position on the x-axis indicating whether the feature has a positive or negative impact on the prediction. Red dots represent higher feature values, or “yes” for categorical variables, while blue dots represent lower values or “no.” For instance, red dots on the positive side suggest that high feature values increase ASI risk, while blue dots on the negative side suggest reduced risk. Features are ranked in order of importance, from top to bottom. Abbreviations: MMSE, Mini-Mental-State-Exam; ADMs, Anti-dementia medication; Past SA, past suicide attempt; LA, living arrangement

Discussion

This study implemented supervised predictive models to improve prediction accuracy beyond the capabilities of traditional statistical methods. While predictive models for suicide risk have been explored [27–30], there have been few that address the unique challenges faced by individuals with cognitive decline [17, 18]. By focusing on this group, our research addresses this gap, providing valuable insights into a high-risk population frequently overlooked in suicide prevention efforts.

The univariate analysis identified several factors associated with ASI, with depressive symptoms at admission and a history of previous suicide attempts emerging as the strongest predictors, aligning with previous research that links depression and suicide risk in older adults [11, 31]. Additionally, those exhibiting better cognitive function, as reflected by higher scores on the MMSE and CDT, or those with a diagnosis of MCI, showed a heightened risk for ASI. Similar findings were reported by Osvath and colleagues [32]. Moreover, previous studies have shown that the duration of cognitive decline significantly influences suicide risk, with the highest risk usually observed in the first six months to a year following a diagnosis [15, 33, 34]. In line with these findings, our study showed that a shorter duration of cognitive decline was associated with an increased risk of ASI. However, it is important to note that a recent meta-analysis did not show a significant association between a recent dementia diagnosis and an increased risk of suicide mortality [35]. If we consider ideation-to-action theories of suicide, most distinguish between suicidal desire and the transition to intent or attempt. Suicidal desire may arise from exposure to psychological or somatic pain, perceived burdensomeness, awareness of deficits, or feelings of entrapment, particularly when protective factors such as connectedness or positive future thinking are weakened. Progression from desire to action requires additional elements, such as increased capability and access to means [5]. In individuals with cognitive decline, these mechanisms may manifest differently across disease stages. Those with mild impairment may retain sufficient awareness of their deficits and perceive themselves as a burden, while still possessing the cognitive capacity to plan and execute suicidal behavior [21, 24]. As cognitive deterioration advances, reduced insight and diminished executive ability may, in turn, lower the likelihood of suicidal action despite persistent emotional distress.

In addition, the use of ADMs was associated with a moderate protective effect against ASI. Similar findings have been documented in prior research, with a Swedish study reporting that ADMs reduced suicide risk, independent of concurrent antidepressant use [36]. One possible explanation for this protective effect is that ADMs may alleviate neuropsychiatric symptoms commonly associated with dementia symptoms that are often linked to increased emotional distress and suicidality [35]. Furthermore, the use of anti-dementia medications may serve as an indicator of a more advanced stage of cognitive decline, which, as discussed above, may inherently reduce suicide risk by limiting the individual’s capacity for planning and executing suicidal behaviors. On the other hand, those with prescribed ADMs are usually regularly appointed to medical care (either primary or specialist psychiatrist/neurologist) and therefore screened for mood/suicidality.

Alongside cognitive factors, psychiatric comorbidities, particularly a history of mood disorders and substance abuse, showed a moderate association with ASI, supporting prior research that highlights a cumulative risk effect of psychiatric disorders on suicidal ideation as well as behavior and even death by suicide among older adults [12, 37, 38]. Regarding physical health, comorbid physical conditions and the presence of pain, were also moderately associated with ASI, consistent with established findings that link chronic illness and pain to increased suicidal ideation in elderly populations [31, 39, 40].

Interestingly, better functional status, as indicated by higher DEMMI scores and stronger performance on the STS test, was linked to a slightly elevated risk of ASI. This finding contrasts with most studies, which have typically associated poorer functional status with increased suicidal ideation [34, 39]. A possible explanation is a transitional phase in which independence is still largely preserved. Early losses in function may heighten distress and perceived burdensomeness, while retained physical/executive capacity could facilitate more elaborated planning. This trend is illustrated in our graphs, where individuals in the early stages of physical decline show the highest incidence of ASI (Figure S1 in Suppl.). Additionally, since our outcome focused on active suicidal ideation (i.e., thoughts with planning or intent), whereas many prior studies combined passive and active ideation, this difference likely contributes to the discrepant findings as well.

On the other hand, sociodemographic factors—such as sex, age, and living arrangement— were not found to be significantly associated with ASI in our univariate analysis, despite their identification as predictors in some other studies [35].

Univariate analysis revealed several variables statistically associated with ASI, though their individual predictive abilities were generally limited. Most predictors exhibited ROC AUC scores between 0.5 and 0.6, indicating discriminatory power only slightly above random guessing (Figure S2 in Suppl.), consistent with findings from a meta-analysis by Franklin and colleagues [41]. Even commonly cited strong risk factors, such as prior suicidal behavior, depression, hopelessness, and male sex, were weak predictors [41]. Notably, in our analysis, depressive symptoms at admission and the duration since cognitive decline onset emerged as the strongest predictors of active suicidal ideation, with ROC AUC scores of 0.70 and 0.67, respectively. However, these findings align with broader research showing that while statistically significant, individual predictors or simple combinations, such as those used in screeners or clinical judgments, often lack strong clinical predictive value, underscoring the need for more complex multivariable models.

To improve the models’ predictions, we developed multivariable models using LR and XGBoost. These models are better equipped to capture (nonlinear) relationships and interactions that univariate analysis may overlook. RFE identified six key predictors for the multivariable models: depressive symptoms at admission, MMSE score, duration of cognitive decline, history of suicide attempts, use of ADMs, and living arrangement (Fig. 1). Furthermore, SHAP values confirmed that depressive symptoms at admission, higher MMSE scores, shorter duration of cognitive decline, and a history of suicide attempts positively contributed to the model’s prediction of ASI (Fig. 3). The use of ADMs was associated with a lower likelihood of ASI, consistent with our univariate findings. As many of these predictors overlap with those identified in the univariate analysis, their associations and potential explanations are not reiterated here. Living arrangement, which showed no significant association with ASI in the univariate analysis, emerged as a relevant predictor in the multivariable models. Individuals residing in nursing homes or living alone showed a higher risk for ASI. This pattern is partly consistent with previous findings linking social isolation and institutional living to increased suicide risk in later life [9, 42]. Perceived burdensomeness has been proposed as a possible mediating factor, although other studies did not confirm such associations [43].

Building on the identified predictors, both XGBoost and LR models demonstrated strong predictive performance, with ROC AUC scores of 0.81 and 0.79, respectively. Importantly, both models showed good capability in identifying the minority positive class, achieving PR AUC scores of 0.55 compared to the baseline of 0.16 for random guessing (Fig. 2). XGBoost marginally outperformed LR in AUC ROC, precision, and F1 score, suggesting it may be the more reliable model for predicting ASI in this population (Table 2). Overall, these results represent a marked improvement over models relying solely on univariately significant predictors, highlighting the robust discriminatory power of machine learning approaches.

This study offers valuable insights into the understanding of ASI of individuals experiencing cognitive decline. However, certain limitations should be acknowledged when interpreting these findings. This study aimed to include a representative sample of patients with cognitive decline, irrespective of their sex, age, or socio-economic background. Given the nature of Slovenia’s health insurance system, the study population is less likely to be affected by socio-economic disparities in access to care. However, the study focused on hospitalized patients. This focus on inpatients may not fully represent those receiving outpatient care, potentially skewing the sample toward individuals with more severe symptoms or those in acute settings. Including data from outpatient visits in future studies could provide a more comprehensive understanding of the broader population and enhance the generalizability of the findings. Secondly, missing data may have introduced bias, as healthcare data are rarely missing completely at random. In this study, missingness was likely related to clinical or cognitive factors, such as symptom severity or assessment feasibility, which could influence both predictor availability and outcome risk. To reduce this bias, k-nearest neighbor imputation was used during variable selection to preserve information across variables. Nonetheless, residual bias from non-random missingness cannot be entirely excluded and should be addressed in future studies. Thirdly, cognitive impairment in participants poses challenges to data accuracy. Individuals with cognitive deficits may struggle to recall or communicate medical histories or personal details reliably. Finally, psychiatric history variables were coded as “no” when not documented, which may have led to slight underestimation of their prevalence. Additional limitations concern the measures used. Apathy is one of the most common neuropsychiatric symptoms in Alzheimer’s disease and mild cognitive impairment. It is conceptually distinct from depression and has been linked to functional decline and amyloid pathology [44]. Systematic assessments of apathy were not available in this retrospective dataset. Future studies should include this construct to clarify its role in suicidality among cognitively impaired individuals. Suicidal ideation was measured using the C-SSRS quick screen rather than a more comprehensive tool such as the Beck Scale for Suicide Ideation, which could have provided greater sensitivity. Cognitive function was assessed with the MMSE, which offers a more limited evaluation of executive and mnestic impairments than the MoCA. Cognitive mechanisms that contribute to suicide risk could threfore not be detected that well, such as impaired decision-making or reduced cognitive control. Finally, the cross-sectional design of this study offers a snapshot of conditions and characteristics at a single point in time without capturing changes over the longer term. A longitudinal design would allow for a more detailed exploration of the progression of suicide risk in individuals with cognitive impairments. This approach could also help identify causative factors, paving the way for more targeted and effective interventions.

Conclusions

In conclusion, our findings support the development of predictive approaches to identify older adults with cognitive decline who may be at increased risk for ASI.

Such models could contribute to earlier recognition and more tailored preventive efforts in psychogeriatric care. With a rapidly aging global population and an increasing prevalence of cognitive impairment, the need for precise, individualized risk assessments is more pressing than ever, particularly in psychogeriatric care settings where older adults face complex, overlapping mental health challenges. Developing EHR-based screening models could assist clinical workflows and help clinicians provide better care for this vulnerable population. Nevertheless, further research is needed to refine these approaches, evaluate their real-world effectiveness, and ensure their readiness for integration into clinical practice.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AD

Alzheimer’s disease

ADMs

Anti-dementia medications (e.g., acetylcholinesterase inhibitors, NMDA receptor antagonists)

ASI

Active suicidal ideation

BMI

Body mass index

C-SSRS

Columbia-Suicide Severity Rating Scale

CDT

Clock Drawing Test

DEMMI

de Morton Mobility Index

Dx

Diagnosis

FRT

Functional Reach Test

ICD-10

International Classification of Diseases, 10th Revision

k-NN

k-nearest neighbours (imputation method)

LA

Living arrangement

LR

Logistic regression

MCI

Mild cognitive impairment

MMSE

Mini-Mental State Examination

NMDA

N-Methyl-D-Aspartate

NPSI

No or passive suicidal ideation

PR AUC

Precision–recall area under the curve

RFE

Recursive feature elimination

ROC AUC

Receiver operating characteristic area under the curve

SA

Suicide attempt

SHAP

SHapley Additive exPlanations

STS

Sit-to-Stand Test

SxS

Symptoms

TUG

Timed Up and Go Test

UD

Unspecified dementia

VD

Vascular dementia

4mWT

4-Meter Walk Test

6MWT

6-Minute Walk Test

10mWT

10-Meter Walk Test

XGBoost

Extreme Gradient Boosting (machine learning method)

Author contributions

Author contributions were as follows: Conceptualization, E.V., A.K.P., P.R.P.; Methodology, E.V., J.R.F., A.K.P., P.R.P.; Formal Analysis, E.V., A.K.P., P.R.P.; Investigation, E.V., A.K.P., P.R.P.; Resources, A.K.P., P.R.P.; Data Curation, E.V., A.K.P., P.R.P.; Writing—Original Draft, E.V., J.R.F., A.K.P., P.R.P.; Writing—Review and Editing, E.V., J.R.F., A.K.P., P.R.P.; Visualization, E.V.; Supervision, A.K.P., P.R.P.; Project Administration, E.V.

Funding

The project was funded from internal institutional funds.

Data availability

Data may be made available from the corresponding author upon reasonable request and with approval from the institutional ethics committee.

Declarations

Ethics approval and consent to participate

The study was approved by the Medical Ethics Committee of the University Psychiatric Clinic Ljubljana. The Committee granted a waiver of informed consent due to the retrospective design. No participant contact or intervention occurred, and all procedures adhered to the 1964 Declaration of Helsinki and its later amendments.

Consent for publication

Not applicable. The study used de-identified retrospective data.

Competing interests

The authors declare no competing interests.