Множественная стратификация риска склероза и прогнозирование расходов на здравоохранение с использованием машинного обучения.

Introduction

Multiple sclerosis (MS) is a chronic, inflammatory, demyelinating disease of the central nervous system, where early diagnosis, treatment, and effective disease monitoring are imperative to reduce long-term patient disability. 1 , 2 MS can vary in presentation and progression, sometimes occurring alongside other chronic conditions, complicating diagnosis and treatment. 3 – 5 Evaluating the efficacy of disease-modifying therapy (DMT) and identifying emergent signs of acute disease relapse is critical to effectively managing patients. 6 – 8 MS patients who experience a relapse are ~3.5 times more costly than their disease peers who did not experience a relapse. 9 – 11 Methods for identifying patients at risk of acute, high-cost events could enable earlier intervention and prevention of avoidable disease exacerbations, improving patient quality of life and reducing costs.

The emergence of large-scale healthcare datasets, including electronic health records and insurance claims data, provides a unique opportunity to explore extensive longitudinal patient healthcare records. 12 , 13 Advanced analytics techniques like machine learning (ML), a subset of artificial intelligence (AI), can analyze large datasets to identify healthcare cost and utilization patterns, potentially aiding in avoiding preventable health events in high-risk patients. 14 – 19 Understanding healthcare utilization patterns preceding acute emergencies, loss of drug efficacy, or general deterioration of health can allow for timely interventions and stratify patients based on emergent risk. Integrating these insights into care management plans could make healthcare more proactive and less reactionary to individual healthcare needs. Individual patients may also change in relative risk over time, making it important to track these changes. 20 Developing ML models to target the most at-risk patients provides an opportunity to focus care delivery individually, promoting better outcomes and reducing healthcare costs.

Previous machine learning applications in healthcare datasets have typically focused on identifying high-cost individuals within the population, regardless of underlying disease or in the setting of acute emergency admissions. These approaches have proven effective in identifying high-spending or at-risk patients at a higher rate than current standard procedures. 21 – 24 ML modeling has also been used to analyze high-cost patients in specific disease categories, demonstrating the potential of this approach to enhance patient prognosis and evaluate healthcare utilization. 25 , 26 Typical limitations of these early studies include the utilization of only a small subset of well-established ML algorithms, such as logistic regression and tree-based modeling, as well as the use of smaller datasets, diminishing the generalizability of trained models. Integration of multiple data sources, such as clinician-initiated administrative data, insurance claims data, patient electronic health records, and patient demographic information, could amplify the predictive signal across healthcare datasets. 27 In the future, the evolution of novel machine learning methods may also offer improved performance compared to classical ML approaches.

This study demonstrates that machine learning models trained on longitudinal commercial claims data can predict which individuals with MS will rank in the top decile of healthcare costs each month. During a held-out evaluation period, these models accurately forecast about 75% of the actual top-decile expenditures and outperformed benchmarks based solely on recent spending patterns. This approach identifies more new high-cost patients and pinpoints those with higher future per member per month costs. These findings suggest that ML-based risk stratification provides a more precise and proactive way to identify rising-risk MS patients compared to retrospective spending rules, enabling earlier, targeted care management and resource allocation.

Methods

Study population & data acquisition

The commercial insurance claims dataset was obtained under a data use license from Blue Health Intelligence. Data access and retrieval were performed on December 1, 2018. In this retrospective study, the dataset comprised a de-identified commercial claims population encompassing over 267,000 individuals across all three-digit zip code areas of a Midwestern state from January 2016 to June 2018. Members included in the study were enrolled with the commercial payer for the entirety of the retrospective study period. The dataset comprised 835/837 healthcare claims from facilities (i.e., hospitals, nursing homes), professional locations (i.e., individual doctors’ offices), and pharmacies. Data fields included Current Procedural Terminology (CPT) codes, Healthcare Common Procedure Coding System (HCPCS) codes, International Classification of Diseases, Tenth Revision (ICD-10) diagnosis codes, National Drug Codes (NDC) pharmacy codes, and de-identified NPI provider codes. Patient demographic information (sex, age, three-digit ZIP code), healthcare use, diagnosis and procedure codes, total utilization, and interaction date were combined to form a base dataset, ensuring consistent formatting across all data points.

Because this retrospective analysis utilized fully de-identified secondary claims data with no direct contact with individuals, the requirement for informed written consent was waived, and institutional review board oversight was not required, in accordance with the U.S. Common Rule (45 CFR 46.104). The study was not registered in a public trial registry for the same reason.

Data preparation & descriptive analysis

The MS patient population was identified as any individual who exhibited an MS ICD-10 diagnosis code (G35) anywhere within their claims record. Individual patient claims were aggregated on a monthly level, creating a unique independent member-month. Quantitative data (healthcare spend, claim counts) were analyzed and aggregated by standard arithmetic calculations, including sums and means. Non-quantitative data, such as alphanumeric codes (ICD-10, CPT, HCPCS codes), were counted during each patient’s monthly interval. All data points were aggregated for each unique individual for a given month during the study timeframe. Patients had varying degrees of healthcare utilization month-to-month throughout the analysis period. Dates and time measurements were only used in the descriptive analysis to categorize the number of days between medical encounters. In the descriptive analysis of the dataset and the ML modeling exercises, the member-month data format was utilized, and each data row represented a month of patient claims data.

Risk analysis

Patients were divided into quartiles based on monthly healthcare spending (the top quartile was defined as the highest spenders). Quartiles were generated by ranking all patients according to their monthly spending and dividing them into four equal groups. Patients were evaluated based on movement across quartiles month to month. Rising-risk patients were defined as those individuals who moved into a higher quartile of spend (greater spending) from one month to the next.

Predictive machine learning modeling

We used a competitive grid search method to compare all developed models and identify the most accurate ones. This was conducted across multiple machine learning modeling families, including regression and classification approaches, to evaluate models for optimal monthly accuracy. These approaches included K-Nearest Neighbors, decision trees, gradient boosting machines, ElasticNet generalized linear models, random forests, support vector machines, and neural networks. The ML experiments were designed to identify high-cost patients at a given timepoint, defined as the individuals comprising the top ten percent of healthcare spenders over the next four months.

The entire dataset was divided into two primary partitions. The first was a model development partition that included training, cross-validation, and holdout datasets (Jan 2016–Sept 2017) used to train and measure model performance. The second partition contained an evaluation dataset (Nov 2017–Feb 2018) to assess generalizability across a previously unseen dataset. This evaluation dataset was employed to simulate real-world model deployment and test performance, as shown in Supplementary Fig. 1 . A four-month interval followed the end of the model development partition to further prevent potential training data leakage. Seventy-two regression-based machine learning models and sixty-three classification-based models were assessed for each prediction month within the evaluation dataset. At each monthly checkpoint in the model development partition, a future four-month spending total was calculated for every individual and used as the prediction modeling target. Choosing a four-month window helped represent future risk trends and smooth out variations in monthly healthcare utilization. ML approaches can adjust to variable predictive timeframes, altering future prediction targets according to specific objectives and data availability. In regression-based modeling, the future four-month rolling sum of healthcare spending served as the training prediction target. In binary classification models, the positive class for each month included individuals who fell within the top ten percent of healthcare spenders over the subsequent four months. The initial modeling experiment used the Jan 2016–Jun 2017 period for training (80%) and hold-out (20%) to determine optimal models for evaluation based on Nov 2017 codes in the evaluation dataset. The training, hold-out, and evaluation datasets each contained a consistent, representative selection of patients with regard to sociodemographic attributes such as race, gender, and age. Subsequent models advanced the timeframe by one month while maintaining an eighteen-month training duration for each modeling study (Supplementary Fig. 1 ). The leading regression and binary classification models were selected based on mean absolute error (MAE) and log loss, respectively, calculated from the 5-fold cross-validation (CV) and the hold-out dataset (Supplementary Table 1 ). The optimal models from each evaluation month were derived from the Gradient Boosted Trees Regressor family of machine learning algorithms. An overview of the model architecture is illustrated in Supplementary Fig. 2 , and a TRIPOD-AI checklist is included as Supplementary Data 1 .

Predictions generated by the selected optimal model for each month in the evaluation dataset were compared to actual monthly healthcare spending to assess model performance. Two alternative risk stratification approaches were compared against ML predictions based on historical healthcare spending in the prior, rolling four-month timeframe or spending in the previous month. Historical spending was calculated for each month and stack ranked, providing a proxy for the predictive value of future patient spending based on historical results. Each monthly prediction was an independent performance test. Each method’s potential risk stratification value was assessed by comparing the method to the actual patient healthcare utilization each month.

Healthcare codes driving the most accurate models were identified through feature impact calculations (Supplementary Table 2 ). Feature impact score is a permutation-based calculation where model accuracy is measured as feature values are shuffled in and out across the entire feature set. Overall impact scores were calculated by averaging feature impacts across the most accurate models developed for each of the four prediction months. To ensure broader relevance to the MS population, only features that appeared in more than fifty member months were reported. These most informative features were then compared to a descriptive analysis of the MS population to compare feature importance with general abundance across the patient population.

Statistics and reproducibility

Statistical analysis was performed using GraphPad Prism software. Data distributions were examined to determine normality, informing the use of parametric or non-parametric tests. Significance ( P < 0.05) was computed across actual and predicted values using an unpaired two-tailed t -test with Welch’s correction (for normally distributed data) or a Mann–Whitney U test (for non-normally distributed data). Multiple comparisons were adjusted using the Benjamini-Hochberg procedure.

The analysis utilized a fixed, de-identified commercial claims dataset comprising over 267,000 members, including 631 with multiple sclerosis, from January 2016 to June 2018. The unit of analysis was the member month, and all available member months within this period were included. Model development used a predefined partition with training, five-fold cross-validation, and a hold-out subset from January 2016 to September 2017. Generalizability was evaluated on an unseen partition covering November 2017 to February 2018, separated by a four month gap to avoid data leakage. Replicates were defined at two levels: cross-validation folds with the development partition and four independent monthly checkpoints in the evaluation period, with each monthly prediction considered as an individual performance test. Permutation-based importance scores were calculated for features within the top-performing models and averaged across the four prediction models. To enhance robustness, only features appearing in at least fifty member months were reported.

Results

Descriptive analysis of the MS and non-MS population

The MS cohort ( n = 631) was primarily middle-aged, with 88% between the ages of 40 and 69. The majority of patients were female (71%), with 29% male (Supplementary Table 3 ). Although fewer in number, male patients spent slightly more each month than female patients, averaging $4760 compared to $4250 per member per month (PMPM). Despite accounting for less than 0.3% of the 267,000-member population, MS patients represented more than 2.5% of total plan expenditures, totaling roughly $102 million of the $4 billion aggregate. On a per capita basis, the average spend was $4398 PMPM for MS patients compared to $665 PMPM for members without MS (Table 1 ). Claim counts increased by fewer than threefold, suggesting that higher member costs, rather than the volume of services, drove the increase in spending. Neurology utilization displayed a similar pattern; MS patients accounted for ~10% of all neurology visits, which was 38 times the rate observed in non-MS members, consistent with guideline-recommended follow-up intervals. 28 Monthly spending was also more variable in the MS group, with a standard deviation of $4101 PMPM compared to $1121 PMPM in the broader population (Table 1 ). These findings highlight the significant and fluctuating economic burden of MS, underscoring the importance of proactive risk-stratification strategies.

The MS patient population exhibited a diverse range of comorbidities, including psychological disorders such as anxiety and depression, chronic pain syndromes, and musculoskeletal conditions like lower back pain and cervicalgia (Supplementary Table 4 ). Cardiovascular conditions, notably hypertension and hyperlipidemia, were also prevalent, underscoring additional factors influencing disease management and healthcare utilization. Preventive healthcare encounters were frequently observed, such as malignancy screenings (cervical and colorectal), immunizations, and routine wellness exams. These findings highlight the complex clinical profiles of MS patients and identify opportunities for targeted interventions to improve outcomes and optimize healthcare resources.

Dynamics of monthly healthcare spend

To capture month-to-month shifts in risk, every member’s spending was ranked each month, and the distribution was divided into quartiles. Individuals were tracked to identify movement between quartiles over time (Fig. 1A ). Most members remained in the same quartile from one month to the next, but those who moved up into a higher quartile recorded the highest average PMPM costs. In any given month, this rising-risk group represented only 20–28% of the population, yet consumed 33–46% of total spending. This pattern was consistent across the entire observation window.

We next focused on the extreme tail of the distribution, defining the top ten percent of monthly spenders as the high-risk, high-cost segment (Fig. 1B ). New entrants to this segment consistently exhibited both the greatest total spend and the highest PMPM outlay, reinforcing the need to identify these individuals before their costs increase.

Machine learning model development and internal validation

ML techniques can process large, complex datasets to estimate the likelihood of recurring events. Utilizing a longitudinal claims dataset, we applied a variety of classification and regression models to determine which MS patients would fall into the top ten percent of monthly healthcare spending over the subsequent four months. For each of the four evaluation windows (November 2017 through February 2018), we trained 63 classification and 72 regression models, then assessed their performance using cross-validation folds and a separate hold-out dataset (Supplementary Table 1 ).

Classification models were assessed based on the area under the receiver operating characteristic (AUC-ROC) curve and log loss, while regression models were evaluated using the R 2 coefficient and mean absolute error (MAE). The top classification models achieved AUC-ROC values ranging from 0.92 to 0.93 and a log loss of 0.24, whereas the leading regression models attained R 2 values between 0.49 and 0.55, with an MAE ranging from $6062 to $6115 (Supplementary Table 1 ). The regression approach more accurately reflected actual spending in the hold-out sample. Therefore, these models were chosen for further testing in a simulated real-world deployment.

Prospective evaluation of MS cost prediction against historical benchmarks

To simulate a real-world setting, we retrained the four best-performing regression models using the complete model development dataset, which included both the training and holdout subsets. We then assessed performance on the evaluation partition dataset from November 2017 to February 2018 (Supplementary Fig. 1 ). The model development and evaluation periods were separated by four months to prevent data leakage. At each monthly checkpoint, we analyzed the model’s predicted top ten percent of spenders and compared these results to two simpler benchmarks: one that relied on total spending over the previous four months, while the other used only the spending from the prior month. All monthly predictions were then compared to the actual high-risk, top-spending group for each month.

Overall, the ML approach captured ~75% of the total expenditures by the actual top ten percent (Table 2 ). In contrast, the four-month and one-month historical benchmarks captured only 39–47% and 34–39% of spending, respectively. On a PMPM basis, the ML predicted group averaged over $14,000 in spending, while the four-month and one-month methods identified individuals who spent $8000 and $6800 PMPM, respectively (Table 2 ).

These three strategies revealed both overlapping and distinct groups of high-cost patients. Notably, the ML models excelled at identifying new entrants into the top ten percent category (Fig. 2 ). A total of $1.4 million in spending was associated with the patients predicted by ML to be a part of the newly emerging group, compared to just $187,000 identified by the four-month historical method. Additionally, the patients identified by ML exhibited a higher average PMPM for new entrants ($19,700 vs. $6800).

Accuracy was further evaluated by measuring how many actual spenders in the top ten percent were correctly identified (Fig. 3 ). The ML approach captured 41–48% of these individuals and 52–61% of the top monthly spending, while the four-month approach identified only 19–30% of members and 20–31% of the total cost. The one-month method demonstrated even lower performance, identifying 13–20% of individuals and capturing 15–22% of top monthly spending. Furthermore, even when the ML model predicted patients outside the actual top ten percent, these individuals still incurred substantially higher costs (~$6800 average PMPM) compared to those identified by the historical four-month analysis (~$4300 average PMPM) or last month’s high-risk group (~$4100 average PMPM) (Supplementary Table 5 ).

We also examined spending patterns for inpatient, outpatient, and pharmacy services among patients identified as part of the rising risk group by ML models (Supplementary Table 6 ). On average, these ML-predicted individuals incurred ~$8000 more in inpatient costs per encounter than those predicted by the four-month historical benchmark. Healthcare spending profiles in the ML-flagged patients mirrored that of the actual rising risk group, reinforcing the model’s ability to accurately pinpoint at-risk individuals. Compared to the historical benchmark, the high average cost per inpatient and outpatient interaction within the ML-predicted cohort suggests that unplanned acute utilization occurs within this patient group, providing an opportunity for proactive intervention and potential prevention of emergency events. These findings suggest that ML-based forecasting identifies patients transitioning into the high-cost segment, providing a better chance to act early and improve outcomes.

Analysis of data feature impact in MS ML model predictions

To identify which healthcare events most strongly influenced our MS cost prediction models, we calculated permutation feature importance scores for every healthcare code. Many of the highest-impact codes were associated with disease-modifying therapies (DMTs) (Supplementary Table 2 ). Although DMT codes were highly informative, they were not the sole drivers of high-risk prediction. Among patients the model predicted to enter the top decile of monthly spending, pharmacy costs did not fully account for the spending gap (Supplementary Table 6 ). On average, pharmacy claims in the rising-risk group were $725 less than in the actual high-risk cohort. Similar spending patterns across inpatient, outpatient, and pharmacy services suggest that the model recognizes a broad utilization signature rather than relying solely on DMT use. For example, natalizumab and ocrelizumab infusions made up 35 percent of CPT costs in the predicted high-cost group versus 31 percent in the actual high-cost group, indicating that risk stratification extends beyond just predicting future medication use (Supplementary Table 6 ).

Next, we compared the top 20 codes with the highest feature-impact scores across top-performing models with the most common codes from a simple descriptive analysis (Supplementary Table 2 ). Only half of these influential ML features overlapped with the most frequently observed codes in the dataset, while the remaining codes involved complex hospital or office encounters that indicate high-cost clinical interactions beyond DMT administration. This highlights how ML-based feature-impact analysis can reveal critical risk drivers that conventional descriptive methods might overlook.

Discussion

ML and other AI techniques are increasingly used to transform data-rich sectors, and healthcare is no exception. 29 , 30 By converting longitudinal administrative claims into actionable risk signals, ML can help clinicians and payers anticipate preventable, high-cost events and proactively allocate resources. This proof-of-concept study demonstrates that a competitive ML framework applied to routine commercial claims can identify individuals with MS who are likely to transition into the top ten percent of healthcare spending within four months, capturing more than 75% of future expenditures and significantly outperforming stratification based solely on historical spending.

Although individuals with MS represented <0.3% of the study population, they accounted for >2.5% of total healthcare spending, consistent with the disproportionate economic burden reported previously. 31 , 32 Importantly, the membership of the top-spending patient cohort varied from month to month, with rising-risk patients contributing the highest costs. Early identification of these individuals provides an opportunity to engage care teams before acute relapses, therapy failures, or avoidable admissions occur. 33 – 37 In practice, predicted patients could be prioritized for interventions such as expedited neurology reviews, care-coordination outreach, telehealth check-ins, or medication optimization, which are approaches that MS specialists regard as both feasible and impactful. 37

Feature impact analysis confirmed that pharmacy claims for DMTs were valuable model inputs, reflecting their well-documented contribution to overall MS costs. 38 , 39 However, the models did not rely solely on DMT codes to flag high-cost patients. Among the rising-risk group identified by the models, the average pharmacy cost per claim was lower than in comparable MS patients, while inpatient and outpatient costs were similar (Supplementary Table 6 ). This parallel cost pattern suggests that models are capturing broader healthcare use trends that lead to cost surges rather than simply identifying patients on expensive medications. High-impact features also included codes for complex hospital encounters, imaging, rehabilitation, and management of comorbidities like hypertension and depression, which are findings consistent with evidence that multimorbidity influences MS outcomes and spending trajectories. 5 , 40 , 41

A common critique of ML in healthcare is its perceived “black-box” nature. 42 To enhance transparency, we provided a flow diagram of the modeling pipeline (Supplementary Fig. 2 ) and a ranked list of influential codes that appeared in ≥50 member-months (Supplementary Table 2 ). These analyses evaluate ML algorithms to help understand how data elements are analyzed and which ones drive predictions, enabling the assessment of whether the outputs are clinically plausible. Future work will extend explainability to the individual patient level, allowing clinicians to view the specific codes contributing to a given patient prediction.

This study examined the potential to anticipate near-term, high-cost MS events using routinely collected claims data; however, several important limitations remain. We did not model long-term disease trajectories, total expenditures, or patient-reported outcomes, making it unclear whether earlier intervention would result in lasting cost reductions or improved health outcomes. Nonetheless, demonstrating that short-term cost surges can be reliably predicted establishes the groundwork for future research into durable management strategies. The analysis also relies on a single-state commercial dataset, so replication across other regions, payers, and care contexts is necessary for broader confirmation. Still, establishing feasibility within one setting is a meaningful first step toward generalizing these methods. Although we required only one ICD-10 G35 code for inclusion, most individuals recorded multiple MS diagnoses over 18 months, suggesting good diagnostic validity. This provides a basis for expanding into more stringent criteria that could further reduce misclassification. Finally, since DMT prescription is ubiquitous and drug pricing largely lies outside clinicians’ control, directing additional attention to non-DMT features in future modeling efforts represents another logical progression for uncovering actionable levers to predict rising-risk patients and contain expenses.

This proof-of-concept study utilized ML modeling to produce accurate predictions for high-risk, high-cost MS patients and has the potential to enhance patient care. A tiered analytics pathway could be implemented, where claims-based ML first identifies rising risks for MS patients, followed by multidisciplinary teams reviewing explainability outputs to determine personalized care. The clinical adoption of this modeling approach will require validation of input claims datasets for accuracy and completeness; any inaccurate or incomplete records will be excluded from model analysis. Prospective studies are necessary to test whether this workflow reduces unplanned admissions, maintains functional status, and offsets implementation costs. Parallel efforts should investigate integration with electronic health records and patient-reported data, extend the prediction horizon, and examine other chronic, high-variability conditions such as inflammatory bowel disease or rheumatoid arthritis. 16 , 43 By iteratively refining the algorithms and surrounding care processes, stakeholders can shift from retrospective cost analysis to anticipatory, value-based management of complex diseases.

Overall, this work demonstrates that routinely collected claims, analyzed through an explainable ML pipeline, can serve as an early-warning system for impending cost escalation in MS. By identifying those likely to incur high expenses and which utilization patterns drive that risk, the model establishes a practical foundation for prospective, multi-center trials that combine real-time risk predictions with targeted care-management interventions. If validated in diverse settings, such a system could be integrated into payer and provider workflows to support value-based contracts, optimize resource allocation, and ultimately reduce the clinical and economic burden of MS.

Supplementary information

Supplementary Material

Description of Additional Supplementary Files

Supplementary Data 1

Acknowledgements

The authors wish to thank Cheryl Sesler, MS, and Mahesh Rao, PhD, for their helpful discussions and for reviewing the manuscript. Neither patients nor members of the public were involved in the design, conduct, reporting, interpretation, or dissemination of this study. This work was supported by Decode Health, Inc.

Author contributions

C.F.S. had full access to all study data and takes responsibility for the integrity of the data and the accuracy of the data analysis. Concept and study design: L.S.W.; J.D.G.; C.F.S. Acquisition, analysis, or interpretation of data: L.S.W.; J.D.G.; C.F.S. Drafting of the manuscript: L.S.W.; J.D.G.; C.F.S. Critical revision of the manuscript for important intellectual content: L.S.W.; J.D.G.; C.F.S. Supervision: C.F.S.

Peer review

Peer review information

Communications Medicine thanks the anonymous reviewers for their contribution to the peer review of this work.

Data availability

The data supporting these findings are available from Blue Health Intelligence under a data use license and are not publicly available. Researchers can request access by contacting Blue Health Intelligence at info@bluehealthintelligence.com. The authors accessed the dataset through the same licensing process available to other researchers. Associated data results are available from the corresponding author upon reasonable request, subject to license terms.

Code availability

Code related to this manuscript can be obtained from the corresponding author upon reasonable request, subject to further confidentiality and licensing agreements, with the explicit restriction that third-party software cannot be published due to contractual obligations.

Competing interests

L.S.W., J.D.G. and C.F.S. are employed by and own stock in Decode Health. The authors declare no non-financial interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

The online version contains supplementary material available at 10.1038/s43856-025-01229-3.

References

1. Tur, C. et al. Assessing treatment outcomes in multiple sclerosis trials and in the clinical setting. Nat. Rev. Neurol. 14 , 75–93 (2018).

2. Berkovich, R. Treatment of acute relapses in multiple sclerosis. Neurotherapeutics 10 , 97–105 (2013).

3. Brownlee, W. J., Hardy, T. A., Fazekas, F. & Miller, D. H. Diagnosis of multiple sclerosis: progress and challenges. Lancet 389 , 1336–1346 (2017).

4. Thompson, A. J. et al. Diagnosis of multiple sclerosis: 2017 revisions of the McDonald criteria. Lancet Neurol. 17 , 162–173 (2018).

5. Magyari, M. & Sorensen, P. S. Comorbidity in multiple sclerosis. Front. Neurol. 11 , 851 (2020).

6. Marriott, J. J., Ekuma, O., Fransoo, R. & Marrie, R. A. Switching to second line MS disease-modifying therapies is associated with decreased relapse rate. Front. Neurol. 14 , 1243589 (2023).

7. Kuhlmann, T. et al. Multiple sclerosis progression: time for a new mechanism-driven framework. Lancet Neurol. 22 , 78–88 (2023).

8. Mankinen, P. et al. Cost assessment modelling of treatments for highly active relapsing multiple sclerosis. Adv. Ther. 37 , 800–818 (2020).

9. Soelberg Sorensen, P. et al. The multiple sclerosis care unit. Mult. Scler. 25 , 627–636 (2019).

10. Parisé, H. et al. Direct and indirect cost burden associated with multiple sclerosis relapses: excess costs of persons with MS and their spouse caregivers. J. Neurol. Sci. 330 , 71–77 (2013).

11. Schauf, M., Chinthapatla, H., Dimri, S., Li, E. & Hartung, D. M. Economic burden of multiple sclerosis in the United States: a systematic literature review. J. Manag Care Spec. Pharm. 29 , 1354–1368 (2023).

12. Pastorino, R. et al. Benefits and challenges of Big Data in healthcare: an overview of the European initiatives. Eur. J. Public Health 29 , 23–27 (2019).

13. Alberto, I. R. I. et al. The impact of commercial health datasets on medical research and health-care algorithms. Lancet Digit Health 5 , e288–e294 (2023).

14. Ricci, B. C., Sachs, J., Dobbertin, K., Khan, F. & Dorr, D. A. Risk stratification in primary care: value-based contributions of provider adjudication. J. Gen. Intern. Med. 37 , 601–607 (2022).

15. Raghupathi, W. & Raghupathi, V. Big data analytics in healthcare: promise and potential. Health Inf. Sci. Syst. 2 , 3 (2014).

16. Tan, S. S., Gao, G. & Koch, S. Big data and analytics in healthcare. Methods Inf. Med. 54 , 546–547 (2015).

17. Zhang, A., Xing, L., Zou, J. & Wu, J. C. Shifting machine learning for healthcare from development to deployment and from models to data. Nat. Biomed. Eng. 6 , 1330–1345 (2022).

18. Alowais, S. A. et al. Revolutionizing healthcare: the role of artificial intelligence in clinical practice. BMC Med. Educ. 23 , 689 (2023).

19. Kumar, Y., Koul, A., Singla, R. & Ijaz, M. F. Artificial intelligence in disease diagnosis: a systematic literature review, synthesizing framework and future research agenda. J. Ambient Intell. Humaniz Comput 14 , 8459–8486 (2023).

20. Reddy, A. et al. Risk stratification methods and provision of care management services in comprehensive primary care initiative practices. Ann. Fam. Med. 15 , 451–454 (2017).

21. Osawa, I., Goto, T., Yamamoto, Y. & Tsugawa, Y. Machine-learning-based prediction models for high-need high-cost patients using nationwide clinical and claims data. NPJ Digit Med. 3 , 148 (2020).

22. Langenberger, B., Schulte, T. & Groene, O. The application of machine learning to predict high-cost patients: A performance-comparison of different models using healthcare claims data. PLoS One 18 , e0279540 (2023).

23. Tamang, S. et al. Predicting patient ‘cost blooms’ in Denmark: a longitudinal population-based study. BMJ Open 7 , e011580 (2017).

24. Snooks, H. et al. Predictive Risk Stratification Model: a Randomised Stepped-Wedge Trial In Primary Care (PRISMATIC). (NIHR Journals Library, 2018).

25. Nghiem, N., Atkinson, J., Nguyen, B. P., Tran-Duy, A. & Wilson, N. Predicting high health-cost users among people with cardiovascular disease using machine learning and nationwide linked social administrative datasets. Health Econ. Rev. 13 , 9 (2023).

26. Suzuki, S. et al. Comparison of risk models for mortality and cardiovascular events between machine learning and conventional logistic regression analysis. PLoS One 14 , e0221911 (2019).

27. Beaulieu-Jones, B. K. et al. Machine learning for patient risk stratification: standing on, or looking over, the shoulders of clinicians?. NPJ Digit Med. 4 , 62 (2021).

28. Wattjes, M. P. et al. 2021 MAGNIMS-CMSC-NAIMS consensus recommendations on the use of MRI in patients with multiple sclerosis. Lancet Neurol. 20 , 653–670 (2021).

29. Doupe, P., Faghmous, J. & Basu, S. Machine learning for health services researchers. Value Health 22 , 808–815 (2019).

30. Goh, K. H. et al. Artificial intelligence in sepsis early prediction and diagnosis using unstructured data in healthcare. Nat. Commun. 12 , 711 (2021).

31. Opara, J. A., Jaracz, K. & Brola, W. Quality of life in multiple sclerosis. J. Med. Life 3 , 352–358 (2010).

32. Hosseini, Z., Homayuni, A. & Etemadifar, M. Barriers to quality of life in patients with multiple sclerosis: a qualitative study. BMC Neurol. 22 , 174 (2022).

33. Flachenecker, P. & Rieckmann, P. Early intervention in multiple sclerosis: better outcomes for patients and society?. Drugs 63 , 1525–1533 (2003).

34. (UK) NCGC. Multiple Sclerosis: Management of Multiple Sclerosis in Primary and Secondary Care (NCGC, 2014).

35. Nazareth, T. A. et al. Relapse prevalence, symptoms, and health care engagement: patient insights from the Multiple Sclerosis in America 2017 survey. Mult. Scler. Relat. Disord. 26 , 219–234 (2018).

36. Ziemssen, T. et al. Optimizing treatment success in multiple sclerosis. J. Neurol. 263 , 1053–1065 (2016).

37. Wenk, J., Voigt, I., Inojosa, H., Schlieter, H. & Ziemssen, T. Building digital patient pathways for the management and treatment of multiple sclerosis. Front. Immunol. 15 , 1356436 (2024).

38. Khakban, A. et al. Direct health care costs associated with multiple sclerosis: a population-based cohort study in British Columbia, Canada, 2001-2020. Neurology 100 , e899–e910 (2023).

39. Bebo, B. et al. The economic burden of multiple sclerosis in the United States: estimate of direct and indirect costs. Neurology 98 , e1810–e1817 (2022).

40. Ponzio, M. et al. The economic impact of comorbidity in multiple sclerosis. Neurol. Sci. 44 , 999–1008 (2023).

41. Bütepage, G. et al. Cost-of-illness trajectories among people with multiple sclerosis by comorbidity: a register-based prospective study in Sweden. Mult. Scler. J. Exp. Transl. Clin. 6 , 2055217320968597 (2020).

42. Chadaga, K. et al. Explainable artificial intelligence approaches for COVID-19 prognosis prediction using clinical markers. Sci. Rep. 14 , 1783 (2024).

43. Allen-Philbey, K. et al. Can we improve the monitoring of people with multiple sclerosis using simple tools, data sharing, and patient engagement?. Front Neurol. 11 , 464 (2020).

Associated Data

Supplementary Materials

Supplementary Material

Description of Additional Supplementary Files

Supplementary Data 1

Data Availability Statement

The data supporting these findings are available from Blue Health Intelligence under a data use license and are not publicly available. Researchers can request access by contacting Blue Health Intelligence at info@bluehealthintelligence.com. The authors accessed the dataset through the same licensing process available to other researchers. Associated data results are available from the corresponding author upon reasonable request, subject to license terms.

Code related to this manuscript can be obtained from the corresponding author upon reasonable request, subject to further confidentiality and licensing agreements, with the explicit restriction that third-party software cannot be published due to contractual obligations.