The MILEPOST Framework for Integrating mHealth Into Community-Driven Interventions: Lessons Learned From a Human-Centered Design Approach in Rural Uganda.

According to the 2021 Global Burden of Disease study, stroke remains the second leading cause of death worldwide []. It is estimated that the number of stroke-related deaths will increase by 50% between 2020 and 2050, with this health burden predominantly borne by low- and middle-income countries []. In China, the economic burden of stroke is particularly heavy; in 2018, its direct treatment costs reached $58.6 billion, of which hospitalization expenses accounted for 79.03% []. Notably, the consumption of medical resources typically exhibits a significantly skewed distribution, with disproportionate medical expenditures often concentrated within a small group of high-cost patients []. This phenomenon of cost concentration is prevalent worldwide [-]. Although interventions, such as interdisciplinary transitional care and complex care management, aim to control costs by optimizing services [-] and have been proven to have certain short-term effects [,], many high-cost patients still face issues of overtreatment or inefficient care []. In this context, implementing rational risk stratification according to patients’ overall disease conditions could potentially address existing shortcomings, ultimately leading to improved cost containment and the optimal allocation of health care resources. Owing to its powerful data processing capacity, machine learning has been extensively used in research concerning health care expenditure prediction. For example, Ma et al [] predicted the average daily costs of patients with psychiatric disorders, Hu et al [] identified the determinants of high costs among patients with breast cancer, and Osawa et al [] developed a predictive model for high-need, high-cost patients. Although previous studies have confirmed the utility of machine learning in predicting medical costs and identifying high-cost patients, it remains necessary to extract latent features closely related to medical expenditures from limited datasets to further enhance model performance. Stroke is typically characterized by a high incidence of multiple comorbidities [], making the effective use of patients’ rich comprehensive diagnostic information crucial. Complex comorbidities not only increase the difficulty of treatment and the risk of mortality but are also closely associated with significant escalations in hospitalization costs [,]. Although existing predictive models generally incorporate comorbidities as key features [-] regarding feature processing, most studies rely on rule-based scoring systems (eg, the Charlson comorbidity index [CCI] and the Elixhauser comorbidity index [ECI]) [,] or simple disease counts. These traditional methods focus solely on the linear accumulation of comorbidities, ignoring the intricate interactions among diseases, which leaves latent information in the data untapped and thereby limits the models’ ability to identify high-cost patients. To overcome the aforementioned methodological limitations, network analysis provides a systematic methodological framework. This network-based analytical paradigm goes beyond traditional simple disease counting; by integrating multiple quantitative indicators such as correlation coefficients, odds ratios (ORs), and the Salton cosine index [], it can effectively evaluate the strength of associations between diseases and subsequently construct comorbidity networks [,]. This topological perspective not only helps identify frequently co-occurring disease clusters but also further reveals the potential interactions between diseases. With the widespread application and standardization of the () codes, it has become possible to construct phenotypic comorbidity networks (PCNs) using hospital discharge data. For example, Xu et al [] improved the prediction accuracy of self-harm behavior by incorporating comorbidity network features, while Hu et al [] effectively predicted patients’ length of stay (LOS) using a multiplex network and a patient similarity network. Notably, Yang et al [] constructed a dual network using the diagnostic records of patients with ischemic heart disease, revealing that the extracted network features significantly outperformed traditional comorbidity indices, thereby effectively enhancing the model’s ability to identify high-cost patients.