The National Health Insurance (NHI) program in Taiwan was initiated in 1995, with the aim of providing mandatory health coverage to the entire Taiwanese population. By the end of 2010, the NHI had a coverage rate of about 99.6%, equivalent to 23 million residents [10]. Over these years, the program leads to an extensive and diverse dataset, which includes detailed information on healthcare utilization, medical diagnoses (coded with the International Classification of Diseases, Ninth Revision, Clinical Modification [ICD-9-CM] during the timespan of this study), treatments, and prescriptions, offering invaluable resources for epidemiological research, healthcare policy analysis, and health economics studies [11]. Its robust data allows for longitudinal studies and facilitates the exploration of health trends, outcomes, and disparities within the population.

We included patients aged 65 or older who had a new dementia diagnosis (ICD-9-CM codes 290.0-290.4, 331.0-331.2, 294.1) from board-certified neurologists or psychiatrists at least twice between January 1, 2002 and December 31, 2009. “Newly diagnosed” indicates no dementia diagnosis prior to January 1, 2002. Patients who returned to the outpatient department with the same new diagnosis were also included. We followed the study cohort until the end of the study (December 31, 2013, or death, whichever occurred first), and collected data including serum evaluations (including a complete blood count and biochemistries, iron, thyroid hormone, vitamin B12, folate, and syphilis), psychological examinations, and brain imaging (computed tomography or magnetic resonance imaging), which is a common procedure in Taiwan when diagnosing dementia. Mortality data was identified from claims records or the registry of catastrophic illnesses [12]. Cases with incomplete records or potential confounders were excluded to ensure a well-defined, high-quality dataset. To validate the data, experienced neurologists and psychiatrists held a meeting to agree on diagnostic variables. The detailed patient selection process and exclusion criteria are illustrated in flowchart (Figure 1). After identifying the study cohort, we randomly allocated 80% of the subjects to a derivation cohort for algorithm training, 10% to a validation sample for external model evaluation, and another 10% to a test sample for final model performance assessment (Figure 2).

These variables were collected from the time of inclusion through the end of the study period, reflecting whether each condition occurred at any point during clinical visits or hospitalizations. If a condition existed at any time within this interval, it was recorded as positive. Before model training, data went under a thorough data cleaning process—assessing structure, completeness, and quality; handling missing data; and validating data accuracy. Demographic information included age, sex, and level of urbanization (ranging from level 1 to level 5, with level 1 being the most urbanized region and level 5 being the least urbanized) [13]. Given our machine learning approach, we initially included a broad range of potentially relevant factors. We then tested all mortality-related variables to identify those with the highest predictive value and removed the others to prevent target leakage. During model development, any variable that reduced overall model performance was also excluded.

The process of training an XGBoost model consists of multiple steps. First, the data was divided into an 80% training group, a 10% validation group, and a 10% test group. Then, the model underwent iterative training by constructing decision trees, with each tree aiming to rectify mistakes made by its predecessors. XGBoost model is used due to its superior calibration, faster learning speed, and that it provides a better balance between bias and variance, which prevents overfitting via regularization [14,15]. Throughout the training process, the model’s hyperparameters, such as learning rate, maximum tree depth, and regularization terms, were adjusted to optimize performance.

Our main goal was to predict the 5-year survival probability, leading us to set the objective function to “binary” for this binary classification task. During the training process, we carefully tuned the hyperparameters to optimize the model’s performance while managing complexity and mitigating overfitting. An initial base score of 0.5 was used to provide a neutral starting point for training.

The subsampling parameters (specifically colsample_bylevel, colsample_bynode, colsample_bytree and subsample) were all set to 1, ensuring that all features were fully utilized at each level, split, and tree. This should maximize the model’s access to information, potentially leading to more accurate and optimal splits, as it considers all variables when making decisions. Additionally, full feature utilization in a comprehensive and representative simplifies the hyperparameter tuning process and ensures deterministic and reproducible results.

The learning rate was set to 0.1, and 1,000 trees were trained sequentially in each iteration (n_estimators=1,000). The maximum tree depth is set to 10, allowing the model to capture complex patterns in the data. Minimum child weight is set to 1, ensuring that nodes are only split when they contain a sufficient number of instances. Regularization terms (reg_alpha=0 and reg_lambda=1) were applied to control overfitting by penalizing large coefficients. The random state is set to 0 so that the results are reproducible (Table 1).

All experimental procedures conformed to the standards set by the latest revision of the Declaration of Helsinki and were approved by the ethics committee of Taipei Veterans General Hospital, Taiwan (protocol code: 2018-07-016AC and date of approval: July 17, 2018). Informed consent was obtained from all subjects involved in the study.

From the initial dataset of 1,000,000 individuals in the Taiwan’s NHI database, and after applying the inclusion and exclusion criteria, a total of 6,556 patients diagnosed with dementia were included in the study cohort, representing an overall prevalence of 1.60%. The prevalence of dementia across various age groups (≥65 years) was distributed as follows: 9.69% in the 65-69 age group, 17.46% in the 70-74 group, 25.95% in the 75-79 group, 25.23% in the 80-84 group, and 15.18% in the 85-89 group, before slightly decreasing to 6.50% in those aged 90 and above. Of the 6,556 patients in this study, 1,777 died within 5 years, yielding an overall 5-year mortality rate of 27.10%.

Within the derivation cohort, those who died within 5 years tended to be older at enrollment, exhibit a higher burden of comorbidities, and experience acute or recent clinical events (e.g., lower respiratory infection, upper gastrointestinal bleeding, and nasogastric tube insertion) in the 3 months before recruitment. They also more frequently received antibiotics, benzodiazepines, and antipsychotics. In contrast, peripheral vascular disease and level of urbanization did not significantly differ between deceased and non-deceased patient. These findings are summarized in Table 2.

These patients were randomly divided into training, validation, and testing groups. For the not deceased category, 3,823 patients were in the training group, with 478 each in the validation and testing groups. For the deceased category, 1,422 patients were included in the training group, and 179 and 178 were allocated to the validation and testing groups, respectively (Figure 2). The average follow-up duration was 6.26 years, ranging from 0.03 to 12 years (standard deviation=5.50, median=6.11). The mean age at enrollment was 79.01 years, with 47.31% of participants being male (2,079 males).

First, we tested all the death-related variables to see which one could achieve the best overall result and found that 5-year death status came out best. As a result, death1y_duration, death1y_status, death2y_duration, death2y_status, death5y_duration, death5y_status, death6m_duration, death6m_status, death_duration, and death_status were omitted. These variables represent survival status at specific time intervals (1 year, 2 years, 5 years, 6 months, and overall status) and were removed because they directly correspond to the defined outcomes and thus cannot serve as independent predictors without causing target leakage. There were several variables that were also removed, which included age_group, habituating_city, and income_level. Age group represented patients’ ages grouped into predefined brackets (e.g., 65-69, 70-74 years, etc.). It was excluded while the precise numerical age variable was being used. habituating_city indicates the city or region where the patient resides, reflecting potential geographic or environmental influences on health outcomes. Income_level categorizes patients’ financial status into distinct tiers. However, these two variables did not improve the model in a positive way despite our effort. Eventually, out of 37 variables, 24 were selected for subsequent model development (Tables 2 and 3).

The predictive model’s performance was evaluated using multiple metrics. It achieved an accuracy of 81.86% and a harmonic mean of precision and recall of 0.66. Precision was 0.68, and recall was 0.63. The true positive rate and true negative rate were 0.63 and 0.89, respectively (Table 4). The confusion matrix (Figure 3A) reported 424 true negatives, 113 true positives, 52 false positives, and 67 false negatives. The model’s logarithmic loss was 0.61. Additionally, the negative predictive value was 0.86, while the false positive rate and false discovery rate were 0.11 and 0.32, respectively (Table 4).

The model’s performance was evaluated using both the receiver operating characteristic (ROC) curve and the precision-recall (PR) curve. The PR curve demonstrated an area under the curve (AUC) of 0.65, illustrating the trade-off between precision and recall across various classification thresholds (Figure 3B). The ROC curve achieved an AUC of 0.81, indicating its ability to distinguish between patients who survived and those who did not within 5 years (Figure 3C).

The XGBoost model identified and ranked the importance of 24 predictors based on their feature weights. Nasogastric tube insertion was the most influential predictor (weight=8.69×10^-2), followed by chronic kidney disease (CKD) (weight=6.20×10^-2), cancer (weight=5.17×10^-2), and lower respiratory infection (weight=5.13×10^-2). Other key predictors included urinary tract infection (weight=4.88×10^-2), femoral neck fracture (weight=4.63×10^-2), and peripheral vascular disease (weight=4.32×10^-2). Additional significant predictors were antidepressant use, hypertension, upper gastrointestinal bleeding, and antibiotic prescription, with weights ranging between 4.26×10^-2 and 4.01×10^-2. Other relevant features included congestive heart failure, employment status, myocardial infarction, dyslipidemia, and stroke, among others, contributing to the model’s predictive capability (Table 3).

Our study identified a dementia prevalence of 1.60% (16,055/1,000,000) among the elderly population, which aligns with previous Taiwanese epidemiological data. A 1998 study from southern Taiwan reported a dementia incidence rate of 1.28% [16], and subsequent studies documented prevalence rates of 1.7%-4.3% among the elderly as of 2008 [17]. In our study, the overall 5-year mortality rate for dementia patients was 27.1%. By comparison, a large-scale Taiwanese study of 372,203 newly diagnosed dementia cases (2004-2013) reported an annual mortality rate of approximately 12% during 5 years of follow-up, and found that the prevalence of most comorbidities increased by 65%-150% after a dementia diagnosis [18]. While these findings broadly align with ours, our more stringent inclusion criteria of receiving a diagnosis of dementia twice may explain the higher cumulative mortality rate. A previous study with similar inclusion criteria reported a 38.7% mortality rate over a mean follow-up of 6.26 years [19].

Nasogastric tube insertion turned out to be the most influential predictor, with lower respiratory infection being the fourth. Both risk factors align with prior research showing that nasogastric tube insertion is associated with increased mortality in advanced dementia. Tube feeding may lead to gravitational back-flow, the presence of the tube across the gastric cardia, and infrequent esophageal body contractions [20]. Previous studies have also suggested that artificial feeding does not necessarily improve nutritional parameters or avoid nutritional worsening in patients with dementia [21-23]. In a prospective study of 67 community-based patients with advanced dementia (Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Functional Assessment Staging Tool stage ≥7A), pneumonia was more common in the patients with a nasogastric tube, as has also been found by others [24]. Cox proportional hazards analysis further revealed that pneumonia, nasogastric tube insertion, and hypoalbuminemia (albumin <3.5 g/dL) were independent predictors of mortality [25].

A previous study with similar population with our research found that patients with CKD are at an elevated risk of dementia [26] and inversely, cognitive impairment significantly increases 1-year mortality risk in older patients with CKD [27]. In another large, matched cohort study compared over 242,000 adults with pre-dialysis CKD to similar individuals without, CKD patients had a higher incidence of dementia early on, and worsening CKD was strongly linked to higher all-cause mortality [28]. CKD may elevate mortality risk due to its frequent coexistence with cardiovascular risk factors, or may reflect more severe, subclinical vascular disease [29], which aligns with the findings in our study, with peripheral vascular disease ranking as the sixth risk factor towards dementia mortality. High incidence and mortality of urinary tract infection were also observed in dementia patients, with delayed treatment significantly increasing the risk of mortality [30,31]. Older adults with both cancer and dementia have higher mortality rates compared to those with cancer alone [32,33], and dementia is an independent predictor of 3-year mortality in patients with femoral neck fractures, contributing to increased vulnerability to complications and subsequent mortality [34,35].

This study demonstrates the effectiveness of using the XGBoost regression algorithm to predict the 5-year survival status of dementia patients and highlights the potential of applying AI in clinical settings. Several variables were removed during model development due to target leakage or lack of positive contribution. For instance, “habituating city” did not improve the model likely because Taiwan’s extensive healthcare accessibility and the relatively chronic nature of dementia reduce the impact of rural residence. Additionally, as Taiwan is a moderately developed country with a smaller wealth gap, “income level” also failed to enhance performance. This may be because Taiwan’s NHI program, which had a coverage rate of approximately 99.6% by 2010 [10], subsidizes most medications, minimizing the effect of income differences on health outcomes. In the future, these currently excluded variables could be reconsidered, and additional environmental or genetic factors might further refine the model.

The model’s performance, as presented by the ROC curve, yielded an AUC of 0.81, displaying discriminatory ability between patients at higher and lower risk of mortality. The curve’s significant rise toward the upper left corner suggests a balanced performance between sensitivity and specificity. The incorporation of these critical comorbidities—such as CKD, cancer, urinary tract infection, and hypertension—enables the model to offer insights that can assist healthcare providers in predicting survival outcomes and prioritizing care for high-risk individuals. However, the trade-off reflected in the PR curve, with an AUC of 0.65, highlights the inherent challenge of maintaining high precision as recall increases. This pattern is common in imbalanced datasets, where identifying more positive cases (higher recall) often comes at the cost of a decline in precision [36,37]. Overall, the model enables healthcare providers to more effectively predict survival outcomes and potentially prioritize care for those at greater risk.

In comparison to traditional models, the XGBoost model has the potential to achieve superior predictive performance. A previous study used Cox proportional hazards regression to create a mortality prediction model in community-dwelling older adults with dementia, and achieved an optimism-corrected integrated area under the receiver operating characteristic curve of 0.76, with time-specific AUCs of 0.75 at 5 years [38]. With the robust dataset from Taiwan’s NHI system and the XGBoost algorithm’s ability to handle complex datasets [9], our study achieved an accuracy of 81.86% and an AUC of 0.81, thus showing effectiveness in managing multiple interacting predictors, capturing the relationships between comorbidities and mortality risk in dementia patients.

For models that have achieved higher accuracy and specificity, a study applied a transfer-learning framework, combining data from both older and younger European populations to refine its model performance, and achieved a geometric accuracy of 87%, specificity of 99%, and sensitivity of 76% [39]. However, Western populations often have different genetic predispositions and lifestyle factors, including grief processes, coping mechanisms, and perceptions of stigma compared to Asian populations [40]. Research also revealed heterogeneity in sex-patterns between ethnics [41], incidence of dementia [42], and even usage of assessment tools [43]. These differences underscore the relevance of research tailored to Asian populations.

Moreover, healthcare systems and access to care vary significantly between regions. Taiwan’s NHI system provides comprehensive longitudinal healthcare data for nearly the entire population, which allows for more in-depth tracking of comorbidities and healthcare utilization patterns over time [44]. In contrast, the European model relies on public datasets with a focus on early biomarkers, which may be less representative of the typical clinical picture in Asian populations. These geographic and demographic differences suggest that while the European model performs well in its context, its direct applicability to Asian populations is limited. Our model, specifically developed for a Taiwanese (Asian) cohort, fills this gap by incorporating the region’s unique healthcare and demographic characteristics, making it a more relevant tool for predicting dementia risk and mortality in this population.

The findings from this study have several potential clinical implications for the management of dementia patients. The predictive model developed using the XGBoost algorithm may be integrated into clinical practice to assist healthcare providers in identifying dementia patients who are at higher risk of mortality. By recognizing high-risk individuals, clinicians can prioritize resource allocation, implement timely interventions, and potentially improve survival outcomes. This approach aligns with personalized medicine, where interventions are tailored to an individual’s specific risk profile, thus optimizing care and minimizing unnecessary interventions.

Furthermore, the integration of this model into healthcare systems, such as Taiwan’s NHI, may contribute to more efficient healthcare planning and resource distribution. This model may serve as a decision-support tool for healthcare providers, potentially aiding in planning for palliative care, monitoring, and targeted interventions based on the patient’s risk profile. The predictive capability of the model could also support discussions with patients and their families about prognosis, facilitating more informed decision-making regarding care goals and expectations.

The study also underscores the potential role of machine learning in enhancing clinical decision-making by providing robust, data-driven insights. As healthcare systems continue to embrace AI, predictive models like the one developed in this study may become integral in managing complex, chronic conditions such as dementia, ultimately contributing to improved patient care and outcomes.

This study offers several key strengths. First, we leveraged a large, population-based dataset from Taiwan’s NHI, covering nearly the entire Taiwanese population. This provides insights into the Taiwanese Han population and may be applicable to broader East Asian populations, in contrast to similar studies that rely on American or European datasets [39,45]. Second, we demonstrated the feasibility of applying AI to develop a prediction model in a medical context. By using the XGBoost algorithm, we managed a complex, high-dimensional dataset and achieved better recall than many traditional analytical methods, providing actionable insights for clinicians and policymakers. Third, focusing on a 5-year survival window delivers practical prognostic information that can guide resource allocation and patient care strategies.

Despite these advantages, there are also limitations. First, we did not capture the degree of severity, frequency, or duration of the variables, which may further influence mortality risk. Another limitation of our study is related to handling class imbalance. Although XGBoost provides the scale_pos_weight parameter to adjust for imbalances by increasing the weight of positive class instances, our study had to use a fixed value of 1.00 due to current software limitations, despite the heuristic suggesting a value of approximately 2.69 based on our data. Data augmentation techniques were not employed, as they are less suitable for the structured, administrative dataset used in our analysis. Our data comes from a single healthcare system, which may limit generalizability to populations with different demographic or healthcare infrastructures. Additionally, relying solely on administrative data—rather than more detailed clinical records—can introduce coding errors or missing information. With recent advancements in large language models and AI algorithms, future studies could benefit from incorporating more granular medical records as well as the currently excluded variables, lifestyle factors, social support, etc. The use of cross-sectional data instead of purely retrospective data could also be considered to support real-time mortality assessments. Due to small sample sizes, we were not able to develop subtype-specific models yet, because doing so could risk overfitting. Dementia with Lewy bodies was often coded with 294.1 (dementia in conditions classified elsewhere) under ICD-9-CM clinically due to health insurance policies. Future work with larger, better coded and more granular datasets may enable distinct prediction models for each dementia subtype. Additionally, a normal-variant cohort would help determine whether the model’s predictive metrics are specifically capturing mortality risk within the dementia population or may reveal more general risk factors present in the broader elderly population. In future research, we would like to incorporate a comparison group to evaluate how well these metrics generalize beyond the dementia population.

In conclusion, this study demonstrates the potential of using the XGBoost algorithm to predict the 5-year survival status of dementia patients, achieving high accuracy and identifying key predictors of mortality. While limitations exist, such as the generalizability of the dataset and the use of administrative data, the model’s strengths in accuracy and the insights provided into key clinical factors are noteworthy. The findings suggest that integrating this model into clinical practice could enhance decision-making, support personalized care, and contribute to more efficient healthcare resource allocation. Future research should address the limitations and explore the inclusion of additional predictors to further refine and improve the model. Overall, this study provides a foundation for advancing dementia care using machine learning technologies.