Explainable machine learning models for identifying mild cognitive impairment in older patients with chronic pain

The global older population is projected to double to 1.5 billion by 2050 [1], with China facing significant aging challenges. By 2021, China had 264 million older adults, expected to comprise a quarter of its population by 2025 [2]. Chronic pain is one of the prevalent and costly health problems in older adults that limits mobility and reduces quality of life [3]. Studies have shown that the incidence of chronic pain among older adults is significantly higher than in the general population, about 40-80%, and this proportion rises with age [4]. Given China’s large aging population and high chronic pain prevalence, greater attention is needed for this group.

Mild cognitive impairment (MCI) is a transition state from normal cognitive aging to dementia. It is characterized by a progressive decline in cognitive function that does not meet criteria for the diagnosis of dementia and is the best “window of intervention” for dementia prevention [5]. Older adults are at high risk of developing MCI, and those with chronic pain are at an even higher risk of cognitive impairment, with prevalence exceeding 50% [6]. Moreover, chronic pain and MCI often interact, increasing dementia risk [3]. Fortunately, MCI has the potential to reverse [5], making early detection in older adults with chronic pain critical for healthy aging. The US Preventive Services Task Force also recommends early prevention of MCI and optimizing assessment tools [7].

Common MCI assessment tools like the Mini-Mental State Examination (MMSE), Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE), Montreal Cognitive Assessment Scale (MoCA), and Self-administered Dementia Screening Questionnaire (p-AD8) are valuable, and annual cognitive health assessments are recommended for older adults [8]. However, 29%-76% of people with cognitive impairment fail to be diagnosed promptly, making early prevention of MCI a challenge [9]. These tools are time-consuming and their cut-off values need to be further clarified [5]. Moreover, traditional methods, based on neuropsychological tests, often result in later-stage diagnoses [10]. Prediction models enable earlier identification of high-risk individuals by integrating specific MCI risk factors, thus addressing the limitation of late diagnosis in traditional tools and providing a more comprehensive and targeted assessment. Therefore, developing such models for older patients with chronic pain are upstream approaches to preventing MCI.

Previous studies have explored the complex relationship between chronic pain and MCI, but few have focused on identifying MCI risk in older patients with chronic pain. Although many risk factors for MCI in older adults are worth considering [11], this needs further validation in this specific subgroup. Moreover, it remains unclear which variables have higher predictive value. Although MCI prediction models have been developed for older patients with diabetes and chronic kidney disease, etc [12, 13]. , no such models exist for older patients with chronic pain. Besides, most existing models rely on logistic regression methods, and their accuracy and practicality require further optimization.

Machine learning has shown its strong potential in identifying MCI risk. Unlike traditional methods, it can uncover hidden relationships between risk factors and highlighting the most predictive ones [14]. However, the lack of clarity about the reasons behind machine learning decisions can hinder trust and usability among nurses. To address this, the Shapley Additive Explanations (SHAP) method can show how risk factor influences the model’s identification of MCI risk [10]. Specifically, SHAP method highlights the most important variables for MCI risk and explain how a single variable, such as pain level or age, or all included variables, affect MCI risk for each individual. SHAP makes it easier for nurses to use machine learning models in clinical settings by explaining the results of MCI risk and variables analyzed by the model in a way that nurses can understand, helping them better assess MCI risk and develop personalized care plans that target specific risk factors early, thereby improving patient outcomes [10].

Therefore, this study aims to determine MCI risk factors in older adults with chronic pain and develop explainable machine learning models to identify MCI risk. The findings will help nurses detect high-risk patients early and develop personalized care strategies to reduce MCI incidence.

This study aimed to enhance the cognitive health of older patients with chronic pain by identifying risk factors for MCI and developing explainable machine learning models for MCI risk assessment. We determined 7 variables significantly associated with MCI: pain level, age, depression, education level, pain duration, sleep quality, number of pain sites. The XGBoost model performed best, with pain level, age, and depression identified as the three most important variables. The prediction model developed in this study offers nurses with a practical tool for early and individualized MCI risk assessment, and identified risk factors can serve as the focus of targeted assessment and intervention in nursing practice.

This study showed that the incidence of MCI in older patients with chronic pain was 27.6%, higher than the global average for the older population (15.5%) but lower than the 54.4% reported by Moriya et al. in chronic pain clinics [3, 26]. This may stem from differences in sample size and assessment methods. Nevertheless, cognitive issues in this population warrant greater attention. On the one hand, healthcare providers pay significantly insufficient attention to their cognitive impairment, with actual clinical diagnosis rates below 5% [6]. On the other hand, both chronic pain and MCI can be mitigated by intervention [27], emphasizing the necessity of early identification of high-risk patients for developing care strategies.

Consistent with previous findings [11], this study confirms a positive association between age and MCI risk. The SHAP feature importance plot further identified age as the second most important variable for MCI risk. This is not surprising, as aging is closely linked to numerous risk factors. Advanced age exacerbates the interaction between chronic pain, white matter hyperintensities (WMH), and hippocampal atrophy and increases the likelihood of multiple chronic comorbidities [28]. Aging also often leads to diminished social capabilities, further compounding MCI risk [29]. Therefore, nurses should focus on cognitive screening and follow-up plans tailored to different age groups, while enhancing chronic disease management and psychosocial support. This study also showed that patients with lower education levels were more likely to develop MCI, consistent with previous research [30]. This may be due to reduced synaptic density in the cerebral cortex and insufficient stimulation of brain cells, which hinder the maintenance of cognitive function [11]. Given that more than 65% of patients in this study had an education level of primary school or below, nurses should implement health education programs, encourage patients to engage in lifelong learning, and guide family members to participate in cognitive health management efforts.

This study identified pain level as the most important variable positively associated with MCI risk, as highlighted by SHAP feature importance analysis. This novel finding can be explained by existing research [31]. Chronic pain has been associated with alterations in the cerebral cortex and gray matter, and elevated cortisol levels, which impair hippocampal synaptic plasticity. Chronic pain, the most common symptom of patients, also triggers a cascade of adverse effects, including depression and activity limitations, often creating a vicious cycle that further deteriorates cognitive function. Notably, we observed an association between the number of pain sites and increased MCI risk, a finding supported by Zhao et al., who reported that chronic pain at multiple sites correlates with more severe cognitive impairment and hippocampal atrophy [32]. In addition, our study found that pain duration was positively associated with MCI risk overall. Interestingly, however, unlike the linear relationship reported in previous study [33], our SHAP dependence plot revealed an inverted “U” curve in the relationship between pain duration and MCI risk, indicating that when pain duration exceeded one year, MCI risk decreased. The “adaptation hypothesis” may offer an explanation for this phenomenon: chronic pain initially induces increased stress and neuroinflammation; however, over time, the brain compensates for sustained pain by utilizing alternative coping mechanisms, leading to psychological adaptation and cognitive resilience, which mitigates cognitive decline [34]. Moreover, as time progresses, the pain and cognitive self-management skills acquired by patients may help reduce MCI risk. Nurses should prioritize comprehensive and regular pain assessments, including assessments of pain level, duration, and number of pain sites. In the early stages of chronic pain, nurses should prioritize cognitive screenings and address potential risk factors for MCI. As pain duration increases, the focus should shift towards maintaining functional independence, enhancing coping strategies, and promoting social engagement to mitigate the long-term cognitive effects of chronic pain.

The SHAP feature importance plot identified depression as the third most important variable associated with MCI risk. This finding aligns with previous research, which has shown that depression contributes to increased cortical amyloid-beta deposition. Moreover, depression and chronic pain often co-occur, and their synergistic effects, compounded by their adverse impact on sleep health, significantly exacerbate cognitive impairment [35]. Our study also confirmed that sleep quality significantly impacts MCI risk, consistent with previous research [36]. Long-term poor sleep quality may reduce acetylcholine and monoamine neurotransmitters levels in the brain, impairing memory and cognitive function [37]. Therefore, regular depression screening and sleep quality monitoring should be integral to cognitive health care strategies. Nurses should promote cognitive behavioral therapy and relaxation techniques to manage depression. Additionally, they should encourage patients to adopt proper sleep hygiene practices, including consistent sleep schedule and minimizing screen exposure before bedtime. However, unlike the findings of Jia et al. [26], this study failed to find a significant association between smoking, alcohol consumption, and MCI. We speculate that the cross-sectional design, which cannot assess temporal causality, along with the small sample size and the rigorous variable selection methods, may have limited our ability to detect these complex relationships. Moreover, hypertension and hyperlipidemia may have mediated their effects on MCI risk, potentially obscuring any direct associations. But given inherent health risks of smoking and alcohol consumption, promoting a healthy lifestyle remains essential.

Given the lack of prediction models for early identification of MCI in older patients with chronic pain, we developed machine learning models to address this gap. Among these, the XGBoost model performed best, with an AUC of 0.925 (95% CI 0.863–0.995). In comparison, Jiang et al.'s traditional diabetes-based MCI prediction model reported an AUC of 0.893 [12], while Huang et al.'s MCI logistic regression model for middle-aged and older adults achieved an AUC of 0.870 [38]. These findings highlight the superiority of machine learning models over traditional approaches, as traditional models often struggle to capture the intricate, nonlinear relationships between MCI and risk factors. XGBoost excels at handling complex, large datasets, effectively capturing both linear and nonlinear associations to enhance MCI prediction accuracy [10], as evidenced by the superior performance observed in our study. Although some traditional models, like Xu et al.'s (AUC 0.897) [13], achieve comparable performance, they can only explain the MCI risk at the overall level and overlook nuanced insights and fail to account for complex interactions between variable like depression and sleep quality, thereby limiting their clinical applicability. By contrast, our SHAP-based machine learning model provides a comprehensive interpretation, offering precise and personalized guidance for clinical practice. For example, using SHAP feature importance plot, we identified pain level, age, and depression as the three most important variables influencing MCI risk in older patients with chronic pain, which had not been reported previously. These results not only highlight priority areas for future MCI research but also offer actionable insights for nursing practice. Nurses can focus MCI prevention efforts on efficient pain management, regular depression screening, and prioritizing care for older adults. SHAP summary, decision, and dependence plots further clarify the individual and combined effects of these critical variables on MCI risk. For example, by identifying patients at high risk due to severe pain or depression, nurses can prioritize these factors and initiate personalized nursing management for pain and psychology early. Moreover, SHAP waterfall plots provide detailed, patient-specific visualizations of MCI risk, enabling nurses to allocate resources more effectively and plan targeted interventions. In contrast, traditional models lack the capability to prioritize risk factors or offer individualized insights, leaving nurses without clear guidance for targeted prevention strategies. In summary, these explainable techniques optimize traditional one-size-fits-all MCI management strategies and demonstrates unique advantages in developing detailed, personalized MCI assessments [25].

This study offers valuable insights for future research and practical applications in nursing. First, identifying MCI risk factors equips nurses with specific indicators for closely monitoring patients and developing targeted interventions. Future studies should examine whether these risk factors apply to specific chronic pain conditions such as trigeminal neuralgia and lumbar disc herniation, thus further refining nursing strategies. Second, as our model was developed using data from the Chinese population, nurses in other regions should validate and adapt the model to ensure its applicability in local settings. Third, we advocate a multidisciplinary approach to the management of MCI and chronic pain. Nurses should work with other healthcare professionals like pain specialists, neurologists, psychologists to develop a comprehensive care plan. Furthermore, this prediction model enables nurses to integrate regular MCI risk screening for older patients with chronic pain into routine assessments. Such proactive approach supports secondary prevention by facilitating early detection and intervention, as well as tertiary prevention by identifying patients at risk of worsening or recurrent MCI, allowing timely adjustments to care plans [26, 39].

To our knowledge, this study is the first to develop machine learning models to identify MCI risk in older patients with chronic pain. The study has several key advantages: First, the integration of LASSO regression and SVM-RFE methods for variable screening ensures more rational variable selection. Most of the selected variables are modifiable, increasing the potential for targeted interventions. Second, we developed 9 representative machine learning models, enhancing the ability to identify MCI. Moreover, we used multiple evaluation metrics to ensure the comprehensive model assessment. Finally, the model is efficient, easy to use, and visualized using the SHAP method, which significantly improves its practical applicability.

In conclusion, this study revealed that pain level, age, depression, education level, pain duration, sleep quality, and number of pain sites were risk factors for MCI in older patients with chronic pain. We successfully developed 9 machine learning models to identify MCI risk, with the XGBoost model performing best. Furthermore, we enhanced model’s clinical utility using the SHAP method. These findings offer a tool for nurses to detect MCI risk early and highlight the value of integrating machine learning techniques into nursing practice. However, this study was based on a cross-sectional design, future research should validate these findings using longitudinal data to confirm the model’s predictive validity for MCI progression.