Development of a mortality prediction nomogram for dementia patients using the MIMIC-IV database

Abstract

The accelerating global population aging underscores dementia as a critical public health challenge. Dementia patients admitted to Intensive Care Units (ICUs) face significantly elevated mortality risks, yet robust predictive models remain scarce. This study aimed to develop a nomogram-based tool for mortality prediction in ICU dementia patients. 2,280 dementia patients were randomly allocated to training (n = 1,596) and validation (n = 684) sets. Cox regression identified mortality predictors, incorporated into a nomogram. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and Decision curve analysis (DCA). Age, race, blood glucose, Oxford Acute Severity of Illness Score (OASIS), antibiotic use, antihypertensive medication, and nephrotoxic drugs were independent predictors of all-cause mortality. The Cox model demonstrated AUCs for predicting 30-day and 90-day mortality in the training cohort were 0.727 and 0.744, respectively. In the validation cohort, corresponding AUCs were 0.684 and 0.706. Calibration curves indicated good agreement between predicted and actual mortality. DCA demonstrated the model’s clinical utility and net benefit. In conclusion, this study developed and validated a nomogram-based tool to estimate mortality risk in ICU patients with dementia. The model demonstrated moderate discriminative ability and clinical utility for risk stratification within the primary study cohort, while also underscoring the challenges of cross-database generalizability. By integrating multiple clinically relevant predictors, the model provides precise individualized risk assessments and offers a novel methodological framework for prognostic evaluation. The predictive capability of the model suggests its potential to facilitate clinical decision-making and improve patient outcomes. 

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-52185-4.

Introduction

Dementia, a clinical syndrome featuring cognitive decline exceeding that expected in normal aging, is associated with a two- to fourfold increase in mortality and imposes a substantial socioeconomic burden¹. By 2050, the annual economic cost of dementia is projected to reach RMB 1.89 trillion in China and RMB 9.12 trillion globally². Therefore, there is an urgent need for early and accurate identification of individuals at the highest risk of death.

Existing prognostic models for patients with dementia rely predominantly on chronic disease characteristics and functional measures, often neglecting acute-phase clinical parameters. Although Zhang et al.³ developed machine-learning-based mortality predictors across various survival intervals, their models failed to incorporate intensive care unit (ICU)-specific biomarkers and physiological indicators, thereby limiting their applicability in critical care settings. External validations by Cronjé⁴ and Wang⁵ further demonstrated model overfitting and suboptimal generalizability, highlighting the necessity for more robust and externally validated modeling strategies.

Recent investigations have integrated acute-phase data to enhance the prognostication of dementia-related outcomes. Hafdi et al.⁶reported that combining CT-derived markers of cerebral atrophy with clinical and cognitive assessments improved predictive accuracy, but such imaging features remain underutilized. Alaka et al. augmented the CAIDE model with resting heart rate for broader applicability⁷. Despite observations by Yorganci et al.⁸that the one-year survival rate following ICU admission in patients with dementia was only 47.5%, existing models remain ill-suited for this acute-on-chronic population. Kim et al. likewise emphasized that high-resolution ICU data continue to be underexploited in outcome prediction⁹.

Importantly, the biological mechanisms underpinning mortality in dementia, such as progressive cognitive impairment, accelerated physiological decline, and altered treatment tolerance, differ substantially from those observed in non-dementia critical illness. Therefore, conventional severity-of-illness scoring systems (e.g., APACHE, SOFA) frequently exhibit calibration drift within this subgroup, leading to misestimation of both low- and high-risk mortality. Such discrepancies may compromise clinical decision-making, including the timing of palliative care initiation¹⁰. The development of a dementia-specific prognostic instrument may therefore enable more accurate individualized risk stratification¹¹and directly inform critical clinical choices¹⁰.

Nomograms, which transform multivariable regression outputs into intuitive graphical calculators, have been extensively utilized in oncology and critical-care survival prediction^12–14. However, no nomogram has yet been designed specifically for acutely ill patients with dementia. The present study addresses this gap by integrating dementia-relevant clinical variables with acute physiological indices to construct the first nomogram capable of providing individualized, bedside estimates of 30-day and 90-day mortality risk in this vulnerable population.

Methods

Source of data

This study utilized the MIMIC database, specifically version IV v2.2, which was released in January 2023. MIMIC-IV v2.2 integrates de-identified clinical data from over 299,712 patients admitted to Beth Israel Deaconess Medical Center in Boston up to 2020. The dataset includes comprehensive information on demographics, vital signs, laboratory results, and clinical metrics, all systematically organized into accessible tables. Unique identifiers enable the linkage of multiple admissions per patient, facilitating longitudinal analyses. Diagnoses are coded according to International Classification of Diseases (ICD) standards, ensuring compatibility for both clinical and research purposes. Access to the database was granted after completing the required training courses and passing relevant certification exams. Notably, this study adhered to ethical guidelines, as the dataset consists entirely of de-identified information, obviating the need for additional informed consent or institutional review board approval. A uniform threshold of 15% for the proportion of missing data was applied as the criterion for variable exclusion; no imputation methods were employed for the remaining variables. Notably, the variables included in the final predictive model exhibited high data completeness, with missingness rates below 5%, thereby minimizing potential bias associated with complete case analysis.

Study population and data extraction

The study population consisted of ICU patients with a confirmed diagnosis of dementia, defined by the presence of at least one ICD-9 or ICD-10 code for dementia during any hospitalization (Table S1). To ensure comprehensive case ascertainment, both coding systems were queried (ICD-9: 290.0–290.4.0.4, 331.0–331.2.0.2; ICD-10: F00-F03, G30-G31). Patient data were retrieved via Structured Query Language (SQL) scripting within PostgreSQL (version 14.0) from the MIMIC-IV v2.2 database. Individuals without a dementia diagnosis were excluded from the analysis. For eligible patients, baseline parameters were collected upon their admission to the ICU. These parameters encompassed the following categories: (1) Demographic information: Age, Weight, Height, Gender, Ethnicity, Insurance status, Marital status; (2) Medication and treatment history: Antibiotics, Antihypertensive drugs, Nephrotoxic drugs, Norepinephrine, Phenylepinephrine, Mechanical ventilation; (3) Comorbidities: Hypertension, Diabetes mellitus, Congestive heart failure, Malignant tumor, Chronic kidney disease (CKD), Acute renal failure, Pneumonia, Stroke, Hyperlipidemia; (4) Laboratory findings: White blood cell count (WBC), Red blood cell count (RBC), Platelet count (PLT), Hemoglobin (HB), Red cell distribution width (RDW), Hematocrit (HCT), Serum sodium, Serum potassium, Serum calcium, Serum chloride, Blood glucose, Anion gap (AG), Prothrombin time (PT), Activated partial thromboplastin time (APTT), International normalized ratio (INR), Serum creatinine (SCr); and (5) Severity scores: Sequential Organ Failure Assessment (SOFA), Acute Physiology Score III (APSIII), Systemic Inflammatory Response Syndrome (SIRS), Simplified Acute Physiology Score II (SAPSII), Oxford Acute Severity of Illness Score (OASIS), Glasgow Coma Scale (GCS), and Charlson Comorbidity Index (CHARLSON). Race was categorized according to the MIMIC-IV data dictionary (White, Black, Asian, Other). White patients served as the reference group, and the independent association between race and mortality was estimated using multivariable Cox proportional hazards regression, adjusted for age and comorbidities. Nephrotoxic drugs were defined as pharmacologic agents with well-documented nephrotoxic potential, including aminoglycosides (e.g., gentamicin), vancomycin, nonsteroidal anti-inflammatory drugs (NSAIDs), radiographic contrast media, and selected chemotherapeutic agents. Overall, 66.8% of patients were exposed to at least one nephrotoxic drug (training cohort: 67.4% vs. validation cohort: 65.5%; P = 0.388). The complete list of Generic Sequence Numbers (GSN) used to identify nephrotoxic drugs is presented in Table S2. Notably, the handling of missing data followed a predefined protocol, ensuring minimal bias in subsequent analyses. The primary outcome measures were mortality rates at 30 days and 90 days post-ICU admission.

Statistical analysis

Continuous variables were summarized as mean ± standard deviation (SD) or median (interquartile range [IQR]) depending on their distribution, which was assessed using the Shapiro-Wilk test. Comparisons between groups were performed using independent t-tests for normally distributed data or Mann-Whitney U tests for non-normally distributed data. Categorical variables were reported as frequencies and percentages, with comparisons conducted using chi-squared (χ²) tests or Fisher’s exact tests as appropriate.

The study population was randomly divided into a training set (70%) and a validation set (30%) to ensure robust model development and evaluation. To assess external generalizability, the eICU-CRD database was designated as a fully independent external test cohort and was not involved in any stage of model development or fitting. Univariate Cox proportional hazards regression was employed to screen potential predictive factors, and variables with statistical significance (p < 0.05) in univariate analysis were further included in multivariate Cox regression. Variables that remained significant in the multivariate model were selected as predictors for constructing the prognostic model and establishing a nomogram.

The prediction performance of the model was evaluated using the area under the receiver operating characteristic curve (AUC). Calibration curves were generated to assess the agreement between predicted and observed probabilities. Decision curve analysis (DCA) was additionally performed to quantify the clinical net benefit of the model. To facilitate clinical application, the final Cox proportional hazards model was implemented as an open-access, interactive web-based calculator (www.shinyapps.io). This platform enables real-time, individualized mortality risk estimation based on the final set of predictor variables identified through the Cox regression analysis. Two-sided P < 0.05 indicated statistical significance. Statistical significance was defined as a two-sided p-value less than 0.05. All analyses were conducted using R software (version 4.3.0), with key packages including survival, pROC, and rms.

Results

Baseline characteristics of the study cohort

A total of 2,280 dementia patients were identified from the MIMIC-IV database, with 55.35% being female and 44.65% being male. These patients were randomly divided into a training cohort (n = 1,596) and a validation cohort (n = 684) to ensure representative sampling. Patient screening for dementia is illustrated in Fig. 1. The baseline demographic and clinical characteristics of both cohorts are summarized in Table 1, with footnotes provided for detailed variable definitions. Comparative analyses between the training and validation groups revealed no statistically significant differences in baseline characteristics (p > 0.05), confirming homogeneity across the two groups.

Fig. 1.

Flowchart of patient selection. Among 3,500 patients initially screened from the MIMIC-IV database, 1,220 were excluded due to incomplete data (n = 800), non-dementia diagnoses (n = 300), and multiple ICU admissions (n = 120). The final cohort included 2,280 patients.

Table 1.

Baseline characteristics of the training and validation cohorts.

Variable	Total (n = 2,280)	Training (n = 1,596)	Validation (n = 684)	P
Age, years	82.61 ± 8.90	82.58 ± 8.95	82.70 ± 8.79	0.753
Weight, kg	69.84 ± 17.65	69.82 ± 17.46	69.87 ± 18.11	0.960
Height, (cm)	166.22 ± 11.28	166.54 ± 11.21	165.45 ± 11.43	0.207
Gender, n (%)				0.110
Female	1,262 (55.35)	866 (54.26)	396 (57.89)
Male	1,018 (44.65)	730 (45.74)	288 (42.11)
Ethnicity, n (%)				0.230
Black	239 (10.48)	172 (10.78)	67 (9.80)
Other/unknown	519 (22.76)	348 (21.80)	171 (25.00)
White	1,522 (66.75)	1,076 (67.42)	446 (65.20)
Insurance, n (%)				0.481
Medicaid	45 (1.97)	30 (1.88)	15 (2.19)
Medicare	1,652 (72.46)	1,147 (71.87)	505 (73.83)
Other	583 (25.57)	419 (26.25)	164 (23.98)
Marital status, n (%)				0.137
Divorced	147 (7.26)	100 (7.04)	47 (7.77)
Married	809 (39.93)	591 (41.59)	218 (36.03)
Single	405 (19.99)	278 (19.56)	127 (20.99)
Widowed	665 (32.82)	452 (31.81)	213 (35.21)
Comorbid disease, n(%)
Hypertension, n (%)	1,141 (50.04)	793 (49.69)	348 (50.88)	0.602
Diabetes, n (%)	728 (31.93)	502 (31.45)	226 (33.04)	0.456
Congestive heart failure, n (%)	698 (30.61)	481 (30.14)	217 (31.73)	0.451
Malignant tumor, n (%)	392 (17.19)	270 (16.92)	122 (17.84)	0.594
CKD, n (%)	531 (23.29)	381 (23.87)	150 (21.93)	0.315
Acute renal failure, n (%)	859 (37.68)	606 (37.97)	253 (36.99)	0.658
Pneumonia, n (%)	791 (34.69)	553 (34.65)	238 (34.80)	0.946
Stroke, n (%)	409 (17.94)	273 (17.11)	136 (19.88)	0.113
Hyperlipidemia, n (%)	925 (40.57)	646 (40.48)	279 (40.79)	0.889
previous medication and treatment history
Antibiotics, n (%)	1,746 (76.58)	1,226 (76.82)	520 (76.02)	0.682
Antihypertensive drugs, n (%)	1,281 (56.18)	880 (55.14)	401 (58.63)	0.124
Nephrotoxic drugs, n (%)	1,523 (66.8)	1,075 (67.36)	448 (65.50)	0.388
Norepinephrine, n (%)	427 (18.73)	289 (18.11)	138 (20.18)	0.246
Phenylephrine, n (%)	233 (10.22)	145 (9.09)	88 (12.87)	0.006
Mechanical ventilation, n (%)	1,590 (69.74)	1,124 (70.43)	466 (68.13)	0.274
Mechanical ventilation hours	33.18 ± 63.12	32.04 ± 58.53	35.84 ± 72.69	0.188
Laboratory indicators, reference range, unit
WBC	12.08 ± 7.22	11.93 ± 7.40	12.43 ± 6.79	0.133
RBC	3.58 ± 0.66	3.57 ± 0.67	3.60 ± 0.65	0.295
PLT	209.34 ± 103.79	207.82 ± 107.45	212.88 ± 94.66	0.297
HB	10.65 ± 1.93	10.64 ± 1.93	10.69 ± 1.92	0.562
RDW	15.00 ± 2.02	15.01 ± 2.06	14.99 ± 1.90	0.854
HCT	32.86 ± 5.75	32.83 ± 5.82	32.92 ± 5.59	0.753
Serum Sodium	140.62 ± 6.15	140.77 ± 6.26	140.28 ± 5.88	0.083
Serum Potassium	4.14 ± 0.57	4.14 ± 0.58	4.12 ± 0.55	0.451
Serum Calcium	8.41 ± 0.73	8.41 ± 0.72	8.41 ± 0.76	0.859
Serum Chlorine	105.73 ± 7.13	105.85 ± 7.15	105.46 ± 7.08	0.237
Blood glucose	142.67 ± 63.26	141.42 ± 61.51	145.60 ± 67.10	0.155
AG, (10–18, mEq/L)	14.80 ± 3.57	14.77 ± 3.48	14.89 ± 3.76	0.458
PT, (9.4–12.5, s)	15.70 ± 8.53	15.71 ± 8.48	15.70 ± 8.66	0.984
APTT, (25–36.5, s)	37.38 ± 20.91	37.23 ± 20.82	37.75 ± 21.16	0.633
INR, (0.9–1.1)	1.44 ± 0.76	1.44 ± 0.81	1.42 ± 0.61	0.460
SCr, (0.5–1.2, mg/dL)	1.34 ± 1.11	1.35 ± 1.15	1.30 ± 0.99	0.334
Disease severity score, points
SOFA score	4.94 ± 2.96	4.96 ± 3.02	4.91 ± 2.82	0.732
APSIII score	49.53 ± 19.20	49.56 ± 19.47	49.44 ± 18.59	0.889
SIRS score	2.42 ± 0.96	2.41 ± 0.96	2.45 ± 0.96	0.344
SAPSII score	42.86 ± 11.96	42.69 ± 12.05	43.25 ± 11.76	0.308
OASIS score	35.04 ± 7.97	35.05 ± 8.03	35.02 ± 7.83	0.934
GCS score	12.34 ± 2.95	12.34 ± 2.96	12.35 ± 2.93	0.961
CHARLSON score	6.88 ± 2.23	6.90 ± 2.26	6.83 ± 2.14	0.465
Outcome-related measures
Length of hospital	8.90 ± 8.31	8.84 ± 7.92	9.03 ± 9.16	0.624
Length of ICU stay	2.93 ± 3.31	2.89 ± 3.11	3.04 ± 3.73	0.309

WBC: White Blood Cell; RBC: Red Blood Cell; PLT: Platelet; HB: Hemoglobin; RDW: Red Cell Distribution Width; HCT: Hematocrit; AG: Anion Gap; PT: Prothrombin Time; APTT: Activated Partial Thromboplastin Time; INR: International Normalized Ratio; SCr: Serum Creatinine)​.

Construction of the nomogram for predicting 30-day and 90-day mortality risk in dementia patients

Initially, univariate Cox proportional hazards regression analyses were performed to conduct preliminary screening of all candidate variables. This procedure identified 15 variables significantly associated with all-cause mortality (P < 0.1), which were subsequently included in further multivariable analyses (Table 2, Supplementary Materials). These variables encompassed patient demographics (e.g., age), laboratory indices (e.g., white blood cell count, serum potassium, blood glucose, anion gap, albumin, RDW), physiological scores (e.g., SOFA, APSIII, SAPSII, OASIS, GCS), treatment-related measures (e.g., norepinephrine use, mechanical ventilation, antibiotic use, antihypertensive drug use, nephrotoxic drug use), and comorbidities (e.g., acute renal failure, pneumonia). The prescreened variables were then entered into a multivariable Cox proportional hazards model, with stepwise selection applied to derive the optimal predictive set. Multicollinearity was assessed using variance inflation factors (VIFs), all of which were below 5, indicating acceptable levels of collinearity. Predictors demonstrating consistent significance and directionality in both the 30-day and 90-day mortality models were retained to construct the final parsimonious model. The multivariable analysis identified seven independent predictors of mortality (Table 3). The risk factors included advanced age (HR = 1.04, 95% CI: 1.02–1.06), elevated blood glucose (HR = 1.01, 95% CI: 1.01–1.01), higher OASIS score (HR = 1.10, 95% CI: 1.08–1.11), and White race (HR = 1.21, 95% CI: 1.05–1.40, compared with other racial groups). Conversely, antibiotic use (HR = 0.52, 95% CI: 0.37–0.74), antihypertensive drug use (HR = 0.68, 95% CI: 0.52–0.90), and nephrotoxic drug use (HR = 0.64, 95% CI: 0.48–0.85) were identified as protective factors associated with improved survival.

Table 2.

Results from univariate and multivariate Cox regression analyses for identifying predictors of mortality in patients diagnosed with dementia.

Variables	Univariate		Multivariate
	HR (95% CI)	P	HR (95% CI)	P
Age (years)	1.05 (1.03 ~ 1.06)	< 0.001	1.03 (1.01 ~ 1.05)	0.007
Weight	0.99 (0.98 ~ 0.99)	0.011	0.99 (0.98 ~ 1.00)	0.156
Height	0.99 (0.97 ~ 1.00)	0.147
WBC	1.02 (1.01 ~ 1.03)	< 0.001	1.01 (1.00 ~ 1.02)	0.177
RBC	1.07 (0.89 ~ 1.29)	0.459
PLT	1.00 (1.00 ~ 1.00)	0.224
HB	1.02 (0.95 ~ 1.08)	0.651
RDW	1.07 (1.01 ~ 1.12)	0.02	1.06 (0.99 ~ 1.13)	0.106
HCT	1.02 (0.99 ~ 1.04)	0.135
Serum Sodium	1.02 (1.01 ~ 1.04)	0.048	1.00 (0.98 ~ 1.03)	0.743
Serum Potassium	1.53 (1.25 ~ 1.87)	< 0.001	0.97 (0.74 ~ 1.28)	0.852
Serum Calcium	0.81 (0.67 ~ 0.97)	0.022	0.97 (0.78 ~ 1.21)	0.782
Serum Chlorine	1.01 (0.99 ~ 1.03)	0.399
Blood glucose	1.01 (1.01 ~ 1.01)	< 0.001	1.01 (1.01 ~ 1.01)	0.046
AG	1.10 (1.06 ~ 1.14)	< 0.001	1.02 (0.97 ~ 1.06)	0.503
PT	1.01 (1.01 ~ 1.02)	0.041	1.00 (0.91 ~ 1.10)	0.959
APTT	1.01 (1.01 ~ 1.01)	< 0.001	1.01 (1.00 ~ 1.01)	0.084
INR	1.13 (1.01 ~ 1.26)	0.043	0.99 (0.35 ~ 2.82)	0.986
SCr	1.08 (0.99 ~ 1.17)	0.077
Mechanical ventilation hours	1.00 (1.00 ~ 1.00)	0.215
SOFA score	1.16 (1.12 ~ 1.20)	< 0.001	1.06 (0.99 ~ 1.14)	0.118
APSIII score	1.03 (1.02 ~ 1.03)	< 0.001	1.00 (0.99 ~ 1.02)	0.794
SAPSII score	1.05 (1.04 ~ 1.05)	< 0.001	1.00 (0.98 ~ 1.03)	0.769
OASIS score	1.09 (1.08 ~ 1.11)	< 0.001	1.06 (1.03 ~ 1.09)	< 0.001
GCS score	0.93 (0.90 ~ 0.97)	< 0.001	1.00 (0.95 ~ 1.06)	0.89
CHARLSON score	1.04 (0.99 ~ 1.10)	0.11
Length of ICU stay	1.00 (0.97 ~ 1.03)	0.912
Gender
Female	1.00 (Reference)
Male	0.93 (0.73 ~ 1.20)	0.585
Ethnicity
Black	1.00 (Reference)		1.00 (Reference)
Other/unknown	2.23 (1.35 ~ 3.68)	0.002	2.25 (1.20 ~ 4.21)	0.011
White	1.48 (0.92 ~ 2.39)	0.105	1.89 (1.04 ~ 3.44)	0.038
Insurance
Medicaid	1.00 (Reference)
Medicare	2.09 (0.66 ~ 6.60)	0.208
Other	1.71 (0.53 ~ 5.50)	0.365
Marital status
Divorced	1.00 (Reference)
Married	1.17 (0.65 ~ 2.10)	0.595
Single	0.78 (0.40 ~ 1.51)	0.462
Widowed	1.57 (0.87 ~ 2.82)	0.135
Hypertension
0	1.00 (Reference)
1	0.84 (0.65 ~ 1.07)	0.161
Diabetes
0	1.00 (Reference)
1	0.80 (0.61 ~ 1.06)	0.118
Congestive heart failure
0	1.00 (Reference)
1	1.09 (0.84 ~ 1.41)	0.539
Malignant tumor
0	1.00 (Reference)		1.00 (Reference)
1	1.46 (1.07 ~ 1.98)	0.016	1.29 (0.87 ~ 1.91)	0.201
CKD
0	1.00 (Reference)
1	1.15 (0.87 ~ 1.51)	0.325
Acute renal failure
0	1.00 (Reference)		1.00 (Reference)
1	1.37 (1.06 ~ 1.75)	0.014	1.00 (0.71 ~ 1.41)	0.993
Pneumonia
0	1.00 (Reference)		1.00 (Reference)
1	1.47 (1.14 ~ 1.88)	0.003	1.28 (0.93 ~ 1.75)	0.131
Stroke
0	1.00 (Reference)
1	1.23 (0.90 ~ 1.69)	0.196
Hyperlipidemia
0	1.00 (Reference)
1	0.77 (0.60 ~ 1.00)	0.053
Antibiotics are used
0	1.00 (Reference)		1.00 (Reference)
1	0.57 (0.42 ~ 0.77)	< 0.001	0.39 (0.26 ~ 0.59)	< 0.001
Antihypertensive drugs
0	1.00 (Reference)		1.00 (Reference)
1	0.56 (0.44 ~ 0.72)	< 0.001	0.58 (0.42 ~ 0.80)	< 0.001
Nephrotoxic drugs
0	1.00 (Reference)		1.00 (Reference)
1	0.61 (0.47 ~ 0.80)	< 0.001	0.56 (0.40 ~ 0.77)	< 0.001
Norepinephrine
0	1.00 (Reference)		1.00 (Reference)
1	2.22 (1.72 ~ 2.88)	< 0.001	1.41 (0.92 ~ 2.16)	0.115
Phenylepinephrine
0	1.00 (Reference)		1.00 (Reference)
1	1.70 (1.22 ~ 2.36)	0.002	0.92 (0.59 ~ 1.43)	0.704
Mechanical ventilation
0	1.00 (Reference)		1.00 (Reference)
1	1.55 (1.12 ~ 2.14)	0.009	1.46 (0.93 ~ 2.28)	0.099

Table 3.

Results from multivariate cox regression analyses for identifying predictors of mortality in patients diagnosed with dementia.

Variables	P	HR (95% CI)
Age (years)	< 0.001	1.04 (1.02 ~ 1.06)
Blood glucose	0.002	1.01 (1.01 ~ 1.01)
OASIS score	< 0.001	1.10 (1.08 ~ 1.11)
Ethnicity
Black		1.00 (Reference)
Other/unknown	0.001	2.51 (1.44 ~ 4.40)
White	0.009	2.05 (1.20 ~ 3.50)
Antibiotics
0		1.00 (Reference)
1	< 0.001	0.52 (0.37 ~ 0.74)
Antihypertensive drugs
0		1.00 (Reference)
1	0.007	0.68 (0.52 ~ 0.90)
Nephrotoxic drugs
0		1.00 (Reference)
1	0.002	0.64 (0.48 ~ 0.85)

HR: Hazard Ratio, CI: Confidence Interval.

In the final predictive model

Each additional year of age was associated with a 4% increase in all-cause mortality risk (HR = 1.04, 95% CI: 1.02–1.06). Each unit increase in blood glucose level corresponded to a 1% increase in all-cause mortality risk (HR = 1.01, 95% CI: 1.01–1.01). Each unit increase in OASIS score was associated with a 10% increase in all-cause mortality risk (HR = 1.10, 95% CI: 1.08–1.11). Conversely, antibiotic use, antihypertensive medication, and nephrotoxic drugs exhibited protective effects: Compared to non-users, patients receiving antibiotics had a 48% lower risk of all-cause mortality (HR = 0.52, 95% CI: 0.37–0.74). Antihypertensive drug users showed a 32% reduction in mortality risk compared to non-users (HR = 0.68, 95% CI: 0.52–0.90). Nephrotoxic drug users demonstrated a 36% lower risk of all-cause mortality compared to non-users (HR = 0.64, 95% CI: 0.48–0.85) (Table 3).

Establishment and evaluation of a nomogram

The nomogram was developed based on the significant predictors identified in the multivariate Cox regression model. The risk score attributed to each factor is presented in Fig. 2, providing a quantifiable measure of mortality risk. Higher scores on this nomogram correspond to increased probabilities of mortality, enabling personalized risk assessments and therapeutic strategies for dementia patients.

Fig. 2.

Nomogram for predicting 30-day and 90-day mortality risk in ICU patients with dementia The nomogram incorporates age, ethnicity, blood glucose, OASIS score, antibiotic use, antihypertensive drugs, and nephrotoxic drugs. Total points correspond to individualized mortality probabilities at 30 and 90 days.

The performance of the nomogram was evaluated using the AUC. In the training cohort, the AUC values for predicting 30-day and 90-day mortality were 0.727 (95% CI: 0.686–0.769) and 0.744 (95% CI: 0.702–0.786), respectively. Similarly, in the validation cohort, the AUC values for predicting 30-day and 90-day mortality were 0.684 (95% CI: 0.621–0.748) and 0.706 (95% CI: 0.642–0.770), respectively (Fig. 3). External validation using the independent EICU-CRD cohort yielded an AUC of 0.53 (95% CI: 0.41–0.65) for predicting 30-day mortality (Fig. 4). These findings suggest that although the model performs well in the MIMIC-IV population, its discriminative ability is limited in the external cohort, likely reflecting differences in data recording practices and patient demographics between the two databases.

Fig. 3.

(a, b). ROC curves of the nomogram for predicting the likelihood of 30-day and 90-day mortality in patients with dementia.

Fig. 4.

ROC curve of the model for predicting 30-day mortality in patients with dementia in the external validation cohort (EICU-CRD).

Clinical use of the nomogram

DCA demonstrated that the nomogram provided a net clinical benefit across a wide range of threshold probabilities (Fig. 5). The results demonstrate that the nomogram provides substantial clinical value by enabling the precise quantification of benefits across various risk thresholds. This highlights the potential of the nomogram to enhance the accuracy and efficiency of clinical decision-making, particularly in identifying patients at high risk of mortality who may benefit from intensified interventions. An online predictive tool was developed for practical application (https://icudementiacalc.shinyapps.io/dynnomapp/), allowing users to adjust variable values via sliders and obtain real-time mortality risk probabilities. The model results were compared with the combined model of common scores. Our study presented the Cox univariate and multivariate regression results of the combined 7 scores, which were compared against commonly used ICU severity scores within the dementia patient cohort (Table 4). Although the discriminative performance of our model was modest, with AUC values ranging from 0.684 to 0.744 across different time points and cohorts, it demonstrated comparable or slightly superior performance relative to several established scoring systems. Additionally, a comprehensive model incorporating all severity scores achieved an AUC of 0.71 (95% CI: 0.67–0.75) for 30-day mortality in the training set and 0.70 (95% CI: 0.64–0.77) in the validation set (Fig. 6). These results suggest its potential utility for mortality risk stratification in this specific patient population, despite the overall moderate predictive accuracy.

Fig. 5.

Decision curve analysis for the prediction of mortality. (a) 30-day mortality in the training cohort. (b) 30-day mortality in the validation cohort. (c) 90-day mortality in the validation cohort.

Table 4.

Univariate and multivariate Cox regression analyses of seven composite scores for predicting mortality in patients with dementia.

Variables	Single-factor model		Multiple model
	P	hR(95%CI)	P	hR(95%CI)
SIRS score
0		1.00(Reference)		1.00(Reference)
1	0.304	2.87(0.38ཞ21.45)	0.438	2.22(0.30ཞ16.57)
2	0.131	4.57(0.64ཞ32.85)	0.340	2.62(0.36ཞ18.88)
3	0.063	6.46(0.90ཞ46.32)	0.372	2.46(0.34ཞ17.82)
4	0.035	8.41(1.16ཞ60.91)	0.363	2.52(0.34ཞ18.49)
SOFA score	< 0.001	1.20(1.16ཞ1.24)	0.019	1.06(1.01ཞ1.12)
GCS score	< 0.001	0.93(0.89ཞ0.96)	0.065	1.04(1.00ཞ1.09)
OASIS score	< 0.001	1.10(1.09ཞ1.12)	< 0.001	1.07(1.05ཞ1.09)
APSIII score	< 0.001	1.03(1.02ཞ1.04)	0.292	1.01(0.99ཞ1.02)
SAPSII score	< 0.001	1.05(1.04ཞ1.06)	0.238	1.01(0.99ཞ1.03)
CHARLSON score	0.092	1.05(0.99ཞ1.10)	0.769	1.01(0.95ཞ1.07)

Fig. 6.

Comparative assessment of the 30-day mortality prediction model incorporating all severity scores.

Discussion

This study identified age, blood glucose, OASIS score, ethnicity, antibiotic use, antihypertensive medications, and nephrotoxic drugs as independent predictors of all-cause mortality in ICU patients with dementia using the MIMIC database. Based on these findings, we developed a user-friendly nomogram, a graphical tool that converts complex statistical models into an intuitive scoring system. The nomogram demonstrated moderate predictive performance, with AUC values of 0.727 and 0.744 for 30-day and 90-day mortality in the training cohort, respectively, indicating its potential to support clinical decision-making. By providing personalized risk assessments and facilitating early interventions, this tool holds promise for improving patient outcomes and guiding treatment strategies in ICU settings. Predictor selection was intentionally focused on baseline variables, including patient demographics, pre-existing comorbidities, and initial laboratory measurements obtained upon ICU admission. These variables were chosen for their immediate availability at the point of care and high data completeness (> 95%), thereby ensuring the model’s practicality and alignment with the clinical need for early risk stratification. While dynamically monitored physiological parameters might provide additional predictive value, they were not incorporated due to high missingness rates (approximately 40–60%) in electronic health records and the complexity of trajectory-based modeling required for their analysis. Future prospective studies are warranted to explore the integration of such time-series data, which may enhance the sensitivity of mortality prediction in acutely ill dementia patients. In our cohort, hyperglycemia was independently associated with increased mortality risk (HR = 1.01, 95% CI: 1.01–1.01, P < 0.001). This finding is consistent with observations by Kim et al.¹⁵in dementia patients with hyperglycemia. The association is further supported by evidence delineating several interrelated pathophysiological mechanisms through which hyperglycemia exacerbates neurodegeneration^16,17. Hyperglycemia promotes β-amyloid deposition and tau hyperphosphorylation¹⁸, contributes to cerebral atrophy and cortical thinning¹⁹, elevates cerebrovascular risk^20,21,and induces systemic metabolic dysregulation^15,22. These pathways are implicated in increased dementia risk in both diabetic and non-diabetic populations^19,23. Although the observed effect size appears modest, its statistical significance underscores its clinical relevance, particularly when through mechanisms involving inflammation and oxidative stress. Diabetic individuals with mild cognitive impairment (MCI) exhibit a 1.7-fold increased risk of progressing to dementia (RR = 1.7, 95% CI: 1.1–2.4) compared to non-diabetic MCI patients²⁴,emphasizing diabetes as a critical factor in dementia progression. This finding is corroborated by the Rotterdam Study²⁵. Hyperglycemia exacerbates neurodegeneration through mechanisms such as microvascular injury, inflammation, oxidative stress, vascular pathologies, and metabolic dysregulation²⁴. These pathways collectively contribute to adverse outcomes in patients with both diabetes and dementia, posing significant challenges for disease management. Early prevention, particularly glycemic control, is crucial to mitigate dementia incidence and progression. Given the bidirectional relationship between hyperglycemia and neurodegeneration, targeted interventions aimed at maintaining optimal blood glucose levels may not only reduce mortality risk but also improve long-term neurological outcomes. Future studies should focus on elucidating the precise biological mechanisms underlying the association between hyperglycemia and neurodegeneration, as well as evaluating the effectiveness of personalized glycemic control strategies in high-risk populations.

OASIS is a validated tool for quantitatively assessing the acute severity of illness in critically ill patients, incorporating comprehensive clinical data such as physiological measurements and laboratory findings to provide an objective estimation of patient outcomes²⁶. In our study, the OASIS score emerged as a robust predictor of mortality (HR = 1.10 per unit increase), consistent with its established role in evaluating acute illness severity²⁷. This finding aligns with prior evidence demonstrating the prognostic utility of OASIS in various patient populations, including dementia patients with infections²⁶.

Further studies have shown a significant correlation between the OASIS score and survival rates among dementia patients following suspected pneumonia, with each incremental point on the OASIS scale corresponding to an increased risk of mortality²⁶. These results underscore the OASIS score as an independent prognostic factor for dementia patients, emphasizing its potential to enhance prognostic accuracy and inform treatment decisions. Specifically, the application of OASIS can facilitate the development of personalized treatment plans and improve patient-centered care in ICU settings.

However, it is important to acknowledge certain limitations of the OASIS score in this context. While it provides valuable insights into disease severity and prognosis, it does not account for specific factors unique to dementia patients, such as cognitive function or comorbid psychiatric conditions. Future research should focus on refining predictive models that incorporate these additional variables to further optimize their clinical utility. Additionally, comparisons with other scoring systems, such as the Sequential Organ Failure Assessment (SOFA) or Acute Physiology and Chronic Health Evaluation (APACHE) scores, could provide valuable insights into the relative performance and applicability of OASIS in different patient populations.

This study identified White race as an independent risk factor for mortality among ICU patients with dementia (HR = 1.21, 95% CI: 1.05–1.40). In contrast, the associations for Black (HR = 0.93) and Asian (HR = 1.05) patients were not statistically significant, a finding consistent with a large-scale cohort study of 8,080,098 dementia patients²⁸. The elevated risk among White patients may be attributed to the cumulative effect of several factors. First, disparities in healthcare access are evident: Black and Hispanic dementia patients are more frequently transferred to emergency departments, whereas White patients undergo more invasive procedures^29,30. This pattern suggests that White patients are more likely to be admitted to the ICU for acute conditions, such as pneumonia, often in the context of pre-existing multi-organ dysfunction. Second, socioeconomic gradients and the uneven distribution of community resources contribute to this disparity. White individuals of lower socioeconomic status may experience delayed medical care due to insufficient community support, resulting in hospitalization only after severe infections develop³¹. This “delayed admission” pattern is significantly associated with pneumonia-related mortality³². Third, differences in infection-inflammatory responses may play a role: White patients demonstrate a higher prevalence of chronic lung diseases and immunosenescence phenotypes³³,potentially leading to exaggerated inflammatory responses and more rapid increases in SOFA scores during suspected pneumonia³⁴. The increased mortality observed in White dementia patients in the ICU appears to be driven by a cascade of “delayed hospital admission → higher baseline severity → preference for invasive treatment,” rather than by intrinsic biological differences alone. Future studies may further elucidate these mechanisms by stratifying subgroups, such as community-acquired versus healthcare-associated pneumonia.

Pneumonia and other infections, particularly urinary tract infections, are common complications in individuals with advanced dementia³⁵. A longitudinal study spanning 18 months demonstrated that approximately 53% of dementia patients residing in nursing homes exhibited fever symptoms, with 41% succumbing to pneumonia³⁶. Existing literature highlights the critical role of antibiotic therapy in improving survival outcomes for dementia patients with pneumonia³⁷. Notably, the association between antibiotic treatment and mortality within the first 10 days is statistically significant (HR = 0.52, 95% CI: 0.31–0.87), underscoring its efficacy in reducing short-term mortality risk³⁸..

Although the overall correlation between antibiotic use and long-term mortality is not significant, early initiation of antibiotic therapy within the 10-day window remains effective in lowering mortality rates³⁸. This aligns with findings from a multicenter cohort study by Givens et al. (2010),³⁷which reported a 30% reduction in 10-day mortality among dementia patients with pneumonia who received antibiotic treatment (HR = 0.52, 95% CI: 0.31–0.87). In contrast, untreated patients exhibit significantly higher observed mortality rates within two weeks compared to those receiving antibiotic therapy³⁹. In clinical practice, physicians in nursing homes, particularly in the Netherlands, consider the prognosis of dementia patients with pneumonia when making decisions regarding antibiotic treatment³⁹.

Dementia predominantly affects the elderly population, with advancing age correlating with an increased susceptibility to cardiovascular diseases, particularly hypertension⁴⁰. Extensive evidence demonstrates a significant association between the use of antihypertensive medications and a reduced risk of dementia-related mortality^39–46. Antihypertensive drugs effectively manage blood pressure levels, thereby mitigating or delaying the progression of microvascular pathology⁴⁷. This has been corroborated by multiple clinical trials, including both observational and interventional studies, which have established the capacity of antihypertensive therapy to slow the advancement of white matter hyperintensities (WMH)⁴⁶..

Mechanistically, antihypertensive drugs improve the prognosis of dementia by enhancing the functional integrity and structural stability of cerebral vessels, inhibiting arteriosclerosis and small vessel disease, and alleviating vascular impairments associated with hypertension, such as microvascular degeneration, compromised blood-brain barrier integrity, and white matter lesion formation^36,48. These enhancements collectively augment cerebral perfusion and oxygenation, contributing to improved cerebral health.

In our predictive model, nephrotoxic medications unexpectedly appeared as a protective factor (HR = 0.64), a finding that likely reflects confounding by indication rather than a direct biological benefit. Potent nephrotoxic agents are often clinically justified for life-threatening infections in patients with limited therapeutic alternatives⁴⁹.

Despite the potential risk of nephrotoxicity, the therapeutic benefits of these medications may outweigh their adverse effects in specific clinical scenarios⁵⁰. Therefore, the use of these drugs likely serves as a surrogate for more aggressive and targeted clinical management in patients deemed salvageable. While these drugs are essential for eradicating severe infections, their administration requires careful monitoring and renal-protective strategies to optimize outcomes⁵¹. Future research should focus on refining predictive models to better incorporate the complex interplay between nephrotoxicity and therapeutic efficacy, ensuring optimal clinical decision-making.

This study has several notable limitations that warrant acknowledgment. First, the retrospective nature and initial reliance on the single-center MIMIC-IV database inherently limit the generalizability of our findings. To mitigate this concern, we performed rigorous external validation using the independent eICU-CRD cohort. The model’s performance in the external dataset (AUC = 0.53) underscores the challenges of generalizing ICU-specific models across different healthcare systems. This result suggests substantial overfitting to the source database and indicates that, although the identified predictors are clinically meaningful, the model may require recalibration or the incorporation of more heterogeneous training data to improve external validity. Prospective, multicenter studies remain warranted to confirm utility across diverse geographic and clinical settings⁵². Second, our decision to exclude variables with > 15% missingness and to perform complete case analysis may have introduced selection bias and reduced statistical power, although the predictors ultimately retained in the final model exhibited high data completeness, partially mitigating this concern. Third, the nomogram developed herein did not incorporate novel biomarkers (e.g., genetic markers, inflammatory mediators) or imaging-derived metrics (e.g., MRI-based volumetric measurements), which could enhance predictive performance⁵³. Fourth, in our analysis, dementia patients were primarily treated as a homogeneous group. Significant heterogeneity across different dementia subtypes or stages of cognitive decline may be overlooked. Emerging research highlights the value of identifying distinct patient subphenotypes to improve prognostic precision⁵⁴. Due to database limitations, important variables such as dementia subtype, baseline cognitive status (e.g., Mini-Mental State Examination scores), and physical activity levels were unavailable^55–57. Although patient heterogeneity was partially approximated using physiological indicators such as the Glasgow Coma Scale (GCS) score, incorporating more granular dementia-specific metrics in future models will be critical for more personalized risk stratification. Fifth, the relatively small sample size and reliance on internal validation methods (e.g., bootstrap cross-validation) may have constrained the robustness of the findings. Subsequent studies should aim for larger, more diverse sample populations and employ external validation techniques (e.g., independent cohort testing) to comprehensively assess the model’s accuracy and generalizability.

Sixth, our model did not consider functional status assessments (e.g., Clinical Frailty Scale) or activities of daily living (ADL) dependency, which are critical components in predicting outcomes in dementia patients⁵⁸. Future research should integrate these variables alongside advanced cognitive evaluations (e.g., Montreal Cognitive Assessment [MoCA]) and functional assessments to refine prognostic tools.

Finally, it is important to emphasize that the associations identified in our nomogram are predictive rather than causal. Future investigations aiming to determine whether specific treatment strategies, such as antibiotic therapy, improve survival in this population should consider adopting a target trial emulation framework⁵⁹. By explicitly designing observational analyses to emulate key features of randomized controlled trials, this approach may help reduce biases such as confounding by indication.

Conclusion

While the current study provides valuable insights into mortality prediction in dementia patients, addressing these limitations through well-designed future studies will further enhance the clinical applicability and reliability of the model.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADL

Activities of daily living

AG

Anion gap

APSIII

Acute physiology score III

APTT

Activated partial thromboplastin time

AUC

Area under the receiver operating characteristic curve

CHARLSON

Charlson comorbidity index

CKD

Chronic kidney disease

COPD

chronic obstructive pulmonary disease

DCA

Decision curve analysis

GCS

Glasgow coma scale

HB

Hemoglobin

HCT

Hematocrit

ICD

International classification of diseases

ICU

Intensive care units

INR

International normalized ratio

IQR

Interquartile range

MADDE

Mortality assessment in dementia

MIMIC

Medical information mart for intensive care

OASIS

Oxford acute severity of illness score

PLT

Platelet

PT

Prothrombin time

RBC

Red blood cell

RDW

Red cell distribution width

SAM

Survival in Alzheimer’s model

SAPSII

Simplified acute physiology score II

SCr

Serum creatinine

SD

Standard deviation

SIRS

Systemic inflammatory response syndrome

SOFA

Sequential organ failure assessment

SQL

Structured query language

WBC

White blood cell

Author contributions

All authors contributed to the study conception and design. Writing - original draft preparation: Qi Deng; Writing - review and editing: Qi Deng, Rong He, Jianli Bai; Conceptualization: Wenluo Zhang; Methodology: Wenluo Zhang; Formal analysis and investigation: Qi Deng, Rong He, Jianli Bai, Ketong Gong, Wenluo Zhang; Resources: Qi Deng; Supervision: Qi Deng, and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Data availability

The datasets generated and/or analysed during the current study are available in the MIMIC-IV database 2.2, https://physionet.org/content/mimiciv/2.2/.

Declarations

Competing interests

The authors declare no competing interests.