Electrophysiological-based automatic subgroups diagnosis of patients with chronic dysimmune polyneuropathies.

Introduction

Chronic Dysimmune Polyneuropathies (CDP) are characterized by an immune-mediated response against peripheral nerve myelin or axons, resulting in a wide range of motor, sensory, and autonomic deficits [ 1 ]. Given the heterogeneity of symptoms and etiopathogenesis, CDP may be further classified into different subgroups. The most common is Chronic Inflammatory Demyelinating Polyneuropathy (CIDP), a condition characterized by demyelination and progressive neurological disability [ 2 ]. It is caused by an aberrant immunological response - cellular and humoral - triggered by the loss of self-tolerance, even in the absence of specific antibodies [ 3 ]. This condition has a prevalence of around 1–9 cases per 100ˈ000 individuals, affecting predominantly male adults [ 4 ]. Clinically, it presents progressive or relapsing, symmetrical motor-predominant and sensory symptoms [ 5 ]. Treatment aims to slow the disease progression through immunosuppressive therapies, plasma exchange, and immunomodulatory drugs [ 6 ]. Another subset encompasses neuropathies associated with Monoclonal Gammopathies (MG), a plasma-cell disorder producing a monoclonal antibody defined as M-protein (M-spike). MG may appear in different forms, including the Immunoglobulin M Monoclonal Gammopathy of Undetermined Significance (IgM-MGUS). This condition adds a layer of complexity to neuropathic diagnosis, particularly when anti-Myelin-Associated Glycoprotein (anti-MAG) antibodies are involved [ 7 ]. This condition includes sensory neuropathy with mild motor symptoms, sensory ataxia, and upper limb tremors [ 8 ]while the anti-MAG-negative form (IgM-MGUS anti-MAG-) typically exhibits milder sensory involvement. About 49-62% of polyneuropathies associated with IgM-MGUS are demyelinating, with varying associated axonal damage and relative slowed nerve conduction [ 9 ]. Treatment options are limited to supportive care and therapies such as rituximab or intravenous immunoglobulins [ 10 ].

Nowadays, the specific diagnosis of CDP (and its subgroups) entails a multidimensional approach involving clinical, laboratory, and neurophysiological exams. Electrodiagnostic (EDX) evaluation is considered the most efficient and valuable tool to investigate these conditions [ 11 ] through sensory and motor Nerve Conduction Studies (NCS) [ 12 ]. A differential diagnosis between CIDP and polyneuropathies associated with IgM-MGUS has a relevant clinical impact, as treatment strategies and related prognosis may differ [ 13 ]. Over the years, different studies have reported the challenge of differentiating these conditions for reaching an accurate diagnosis [ 14 , 15 ]. Currently, the classification is based on the European Federation of Neurological Societies (EFSN) and the Peripheral Nerve Society (PNS) guidelines [ 16 , 17 ]which provide standard criteria in terms of investigation, diagnosis, and treatment. According to the guidelines, the diagnosis can be defined when specific clinical characteristics are present in combination with predefined EDX abnormalities, and/or accompanied by additional diagnostic evaluations. However, subgroups appear to have numerous overlapping features, which may complicate the classification. Moreover, this diagnostic gold standard, mainly based on classical analysis of electrophysiological parameters, may suffer from a lack of clarity [ 18 ]. For instance, the heterogeneity in patients characteristics may not be distinctive enough to clearly differentiate between subconditions [ 19 ]and the lack of classification of complex atypical forms [ 20 ] results in a reported misdiagnosis rate of 45% in CIDP patients [ 21 ]. Indeed, while specific electrophysiological patterns may differ, both conditions share alterations in nerve conduction parameters, underscoring the challenge of accurate differential diagnosis and highlighting the pressing need for standardized stratification procedures to optimize the development of tailored therapeutic strategies [ 22 ].

Advanced signal analysis techniques and properly validated Machine Learning (ML) models based on electrophysiological features did not find wide employment in neuropathies assessment [ 23 , 24 ]. However, the use of these automated algorithms may help in the identification of complex patterns within the data, enhancing the accuracy in the identification of CDP subgroups. Pioneering studies have progressively demonstrated the effectiveness of these methods in assisting clinicians with the evaluation and classification, even discerning intricacies within large datasets. In the study of Alarcón-Narváez et al. [ 25 ]various ML classifiers were tested and compared using merely clinical variables to diagnose different subtypes of Guillain-Barré Syndrome (GBS). Accordingly, Uncini et al. [ 23 ] proved the excellence in classifying between CIDP and no-CIDP using the same model with electrophysiological variables. Others have explored the classification of polyneuropathies through the use of unsupervised algorithms [ 26 , 27 ] using agglomerative hierarchical clustering approaches on the electrophysiological data. However, to our knowledge, no studies have been found regarding the use of ML methods for the distinction between the three different neuropathies: CIDP, polyneuropathy associated with IgM-MGUS anti-MAG+, and CIDP-like polyneuropathy associated with IgM-MGUS anti-MAG-.

The aim of the study is to investigate the potential of ML algorithms based on a comprehensive set of electrophysiological data to enhance the diagnostic accuracy and interpretability in distinguishing among three subgroups of polyneuropathies. Indeed, although subtypes differ in pathogenesis, prognosis, and therapeutic approach, overlapping features often lead to diagnostic uncertainty [ 14 ]. The two main objectives of this work include: (i) the evaluation and comparison of different ML classifiers in distinguishing among the three subgroups, and (ii) the identification and interpretation of the most informative neurophysiological parameters by embedding ML models with explainability techniques. The work aims to provide clinicians with standardized and objective tools and algorithms for the classification of neuropathies, improving patient care and management.

Materials and methods

Study population

This study retrospectively included 67 patients with CDP under the care of the Neuromuscular Units of Careggi University Hospital in Florence between January 2022 and April 2024. Patients were stratified in (i) CIDP, (ii) polyneuropathy associated with IgM-MGUS anti-MAG+, and (iii) CIDP-like polyneuropathy associated with IgM-MGUS anti-MAG-. The diagnosis was conducted based on clinical features and EDX studies according to the EFSN/PNS guidelines [ 16 , 17 ]. More information on the population is provided in Table 1 .

The inclusion criteria were: (i) age over 18 years; (ii) initial EMG examination within 3 years from the onset of neurological symptoms with subsequent diagnosis of the disease. The exclusion criteria were: (i) any myelopathy or central demyelination problem; (ii) assumption of cytotoxic drugs, neoplasia, or infections at the moment of diagnosis; (iii) concomitant pathologies caused by other clinical disturbances; (iv) diagnosis of multiple myeloma, plasmacytoma, POEMS syndrome (Polyneuropathy, Organomegaly, Endocrinopathy, Monoclonal plasma cell disorder, Skin changes), or Multifocal Motor Neuropathy with and without conduction blocks. The study protocol was approved by the Ethical Committee of Area Vasta Centro (CEAVC), Florence (n. 17024, 31/03/20).

Neurophysiological recordings

All patients underwent an EDX evaluation using a Synergy electromyograph (Oxford Instruments, Old Woking, Manor Way, England), which automatically extracted numerical values of amplitudes, conduction velocities, and latencies for each nerve. Bilateral nerve conduction studies were conducted on the upper and lower limbs of all patients, covering at least eight motor nerves (median, ulnar, tibial posterior, common peroneal bilaterally) and six sensory nerves with antidromic technique (median, ulnar, and sural bilaterally). The F-waves were recorded using standard techniques. The skin temperature was maintained above 33 °C in the upper limbs and above 30 °C in the lower limbs. Neurography values were referenced to upper and lower limits of normality, which were calculated from 80 healthy subjects in our laboratories in the study of Scarpino et al. [ 28 ]. Several electrophysiological parameters were included in the analysis, particularly (i) Motor Distal Latency (MDL) and Sensory Distal Latency (SDL), (ii) Motor Conduction Velocity (MCV) and Sensory Conduction Velocity (SCV), (iii) Compound Muscle Action Potential (CMAP) and Sensory Nerve Action Potential (SNAP) amplitudes, (iv) Minimal F-wave latency (Fig. 1 ). Variables were considered separately for the right and left side of the body. Overall, 52 features were included in the analysis.

Statistical analysis and model implementation

A descriptive analysis using the median and the interquartile range was conducted on the features, which consisted of numerical variables. A group-level Kruskal-Wallis test was applied to assess overall differences across the three diagnostic groups. For features showing significant global differences ( p < 0.05), Dunn-Bonferroni post-hoc tests were performed to account for multiple comparisons. The analyses were performed on MATLAB R2022b. For ML model implementation, a numerical imputation was conducted to handle missing data. After the exclusion of variables with an excessive amount of missing data, the remaining missing values were imputed using the median of the corresponding feature within the same group. Different ML algorithms were implemented and optimized using Python libraries Scikit-learn [ 29 ] and Optuna [ 30 ]. Specifically, five models were evaluated: Elastic-Net regularized Logistic Regression [ 31 ] (LR), Support Vector Machine [ 32 ] (SVM), K-Nearest Neighbors [ 33 ] (KNN), Random Forest [ 34 ] (RF), and Decision Tree Classifier [ 35 ] (CART). A nested cross-validation strategy was implemented, ensuring that data used for training, optimization, and testing are always kept separated, avoiding any sort of double dipping. In the outer loop, a Leave-One-Subject-Out (LOSO) strategy was employed for maximizing the training data, holding out all data from one subject as the test set while using the remaining subjects for the training set. Within the inner loop, a five-fold cross-validation loop was implemented for hyperparameters optimization, as a trade-off between computational efficiency and reliable tuning. Models’ performances were assessed by comparing test evaluation metrics across all the investigated classifiers. Furthermore, to enhance model interpretability, an analysis of features’ contribution to classification was performed. The SHapley Additive exPlanations (SHAP) technique was applied to the best-performing model identified in the inner loop. This method computes Shapley values to determine the clinical relevance and contribution of each parameter. Specifically, the Shapley value for a specific feature was determined by calculating the average change in the model’s predictions taking into account all possible ways in which features interact with each other [ 36 ].

Results

All sixty-seven patients were included in the study. However, two electroneurographic variables (namely, F-wave latency in SPE both on the left and right side) were excluded from the analyses for an excessive amount of missing data across all subjects. Significant differences were found between SCV-related features, primarily in distinguishing between CIDP and polyneuropathy associated with IgM-MGUS anti-MAG+. A majority of variables involving F-wave latency and MCV exhibited significance in discriminating between CIDP and polyneuropathy associated with IgM-MGUS anti-MAG-. Particularly, F-wave latency resulted significantly higher in CIDP, whereas MCV significantly increased in polyneuropathy linked to IgM-MGUS anti-MAG-. Amplitude-related features, namely CMAP and SNAP, present significant differences between CIDP and IgM-MGUS anti-MAG + associated polyneuropathies. Finally, the MDL and SDL mostly contributed to classifying between the two subtypes of IgM-MGUS-related polyneuropathies. All significant variables ( p < 0.05) for each binary classification are reported in Supplementary Materials.

The training, validation and test set accuracies for all five models are presented in Table 2 . The results showed that, except for the KNN, all the other classifiers achieved a test set accuracy exceeding 70%. The best model resulted in the RF with a validation accuracy of 86.5% and a resulting test accuracy of 80.6%. The related test-set confusion matrices are reported in Fig. 2 .

The analysis of features’ contribution was conducted on the models with the higher validation accuracy, namely the RF. The results regarding the impact of each parameter on the output of the RF model showed the relevance of F-wave latencies, MDL, SDL and MCV for group discrimination. In Fig. 3 , the first ten parameters with the greatest impact on the model output are reported. Moreover, this analysis was conducted by comparing one class against the other two. The results for the best fifteen parameters are reported in Fig. 4 , where the Shapley values for each patient and feature-wise are shown.

In Table 3 , a summary of all features’ behavior is provided. The analysis emphasized the pathological role of F-wave latency in discriminating CIDP from the other classes. Regarding conduction velocity, a similar behavior was evident for MCV, whereas SCV identified polyneuropathy associated IgM-MGUS anti-MAG+. Instead, in alignment with the common motor-dominant nature of CIDP and sensory-dominant behavior of polyneuropathy associated with IgM-MGUS anti-MAG+, a pathological decrease in MCV for the first group and in SCV for the second one showed relevant discrimination capability. Interestingly, in contrast to a typical reduction, the atypical less impaired value of CMAP emerges as a possible biomarker for polyneuropathy associated with IgM-MGUS anti-MAG- compared to the other two groups. On the other hand, the pathological increase of motor and sensory distal latency appears to help in distinguishing polyneuropathy linked with IgM-MGUS anti-MAG+. The SNAP did not play a significant role in the model. These results primarily referred to the median and ulnar nerves on either side.

Discussions

The diagnosis of CDP is often challenging and requires an integrated approach combining clinical, anamnestic, and instrumental data. Although the multimodal investigation may help in the identification of a specific condition, different studies have highlighted the difficulties in the differentiation due to the presence of overlapping features [ 37 ]. Our study aims to support clinical interpretation by introducing a data-driven automatic classification method based on electrophysiological variables. The integration of advanced analysis may enrich the diagnostic framework providing elements for a faster and more accurate comprehension of different subtypes of polyneuropathies. As observed in different fields, data-driven ML algorithms can support clinical decisions and reduce the uncertainty decision level of the clinical team [ 38 ].

Previous studies qualitatively attempted to identify possible predictive properties of clinical and electrophysiological findings, despite the difficulties in distinguishing between CIDP and polyneuropathy associated with IgM-MGUS [ 7 , 23 ]. Clinically, IgM-MGUS is characterized by more pronounced sensory deficits, particularly affecting the sural nerves, whereas CIDP typically involves both proximal and distal muscle weakness and tends to respond more favorably to immunomodulatory therapies. For what concerns electrophysiological features, polyneuropathies associated with IgM-MGUS presented significantly prolonged DML, lower CMAP amplitudes in lower-limb nerves, and reduced terminal latency indices (TLI) compared to CIDP. Conversely, CIDP patients often show conduction block (CB), prolonged F-wave latencies, and a broader distribution of abnormalities across multiple nerves [ 39 ]. However, again, the presence of many overlapping parameters may lead to a less accurate classification. For this reason, the solely qualitative analysis of electrophysiological and clinical data, even combined with other diagnostic tests, may not reach high accuracies in the identification of CDP subgroups. The use of automated methods offers several advantages, such as a more refined analysis, and the possibility to highlight complex interactions.

In our study, all the compared classification algorithms showed promising results with the RF emerging as the best-performing model, achieving a validation accuracy of 86.5% and a test accuracy of 80.6%. Similar analyses were conducted by Uncini et al. [ 23 ]where the SVM applied to a comparable dataset demonstrated excellence in classifying between CIDP and other types of neuropathies. Although a direct comparison is not possible, since we aimed to discriminate between three subgroups of CDP, our method accuracy outperformed the clinical interpretation accuracy found in Uncini et al. [ 23 ] which ranged from 54.5 to 81.8%. Moreover, our study presented an explainability analysis of the electrophysiological features for the identification of the most informative ones. Others explored the classification of polyneuropathies through the use of unsupervised algorithms [ 26 ] such as the agglomerative hierarchical clustering approach on electrophysiological data to discriminate patients with CIDP between distinct subgroups based on the similarity of features. Interestingly, up to now the current landscape of ML applications in polyneuropathy classification predominantly focuses on CIDP, and there is a notable gap in the literature regarding the use of algorithms for the classification of polyneuropathies associated with IgM-MGUS.

Through a class-specific analysis, interesting results emerged regarding the contribution of the different electrophysiological parameters to the classification process. Indeed, an interpretability analysis was conducted through the application of the SHAP technique, typically used in game-theory approaches, able to assess the weight of each variable on the prediction in a patient-specific manner [ 40 ]. This approach provides an understanding of the functions and inter-relationships among features, thereby enhancing the translational utility of ML models in clinical settings. It enables clinicians to comprehend the factors influencing outcome predictions, thus empowering them to determine the reliability of such solutions and when they can be effectively utilized [ 41 ].

The increase of F-wave latency emerges as a pivotal factor in distinguishing CIDP in agreement with existing literature [ 42 ]. This parameter provides crucial insights into the motor component of peripheral nerve function becoming particularly relevant for a mainly motor-centric disorder such as CIDP. Indeed, according to the literature, although both motor and sensory symptoms may affect patients with CIDP, the ensemble of electrophysiological criteria considered for the diagnosis is based on motor nerves and rarely takes into account the impairment of sensory pathways [ 43 ]. The observed prolongation of F-wave latency suggests demyelination and dysfunction in the motor neurons [ 44 ]aligning with the distinctive features of CIDP which predominantly impacts motor pathways. Conversely, a non-affected F-wave latency in both IgM-MGUS-related polyneuropathies suggests a neuropathic involvement more oriented toward sensory pathways, according to the sensory nature of these types of neuropathies [ 16 ]. This distinctive pattern in F-wave latency may serve as a valuable diagnostic indicator and contributor to the differentiation, reflecting significant neurophysiological differences between CIDP and IgM-MGUS neuropathies. For the same reasons, the impact of other motor-related parameters plays a crucial role in CIDP identification [ 43 ]. In terms of conduction velocity, a decreased value in motor pathways serves as a predictive factor for classifying CIDP whereas an inverted pattern was shown for the two subtypes of IgM-MGUS neuropathies. Conversely, an abnormal value in SCV emerges as a key factor in identifying polyneuropathies associated with IgM-MGUS anti-MAG+, while it appeared less impaired for CIDP and those related to IgM-MGUS anti-MAG-. Similarly, the CMAP plays a role in CIDP identification, further reinforcing the consistency of our results. In agreement with these findings, the EFNS/PNS guidelines for CIDP management uniquely considered motor-related parameters in their EDX criteria [ 17 ]. In further accordance with our findings, Koski et al. [ 45 ] proposed two sets of diagnostic criteria for CIDP classification. One of these included the evaluation of recordable CMAPs and either abnormal MDL, MCV, or F-wave latency. The validation of these criteria leads to an 83% sensitivity and a 97% specificity. This emphasizes the significance of motor pathways in the context of CIDP identification, while the consideration of sensory parameters remains unused from the criteria [ 46 ]. Interestingly, an absence of alteration in CMAP appeared to be relevant for the identification of polyneuropathies associated with IgM-MGUS anti-MAG-, whereas it seemed to be not related to those related to IgM-MGUS anti-MAG+. For what concerns SNAP, this parameter was not involved in the classification of any type of neuropathy. Lastly, abnormal values of both motor and sensory distal latency proved to be fundamental in identifying polyneuropathies linked to IgM-MGUS anti-MAG+. Surprisingly, there was the opposite pattern for distinguishing those related to IgM-MGUS anti-MAG-. The role of distal latency in differentiating between the two IgM-MGUS-related polyneuropathies might be attributed to the specific characteristics of neuropathies associated with these conditions. Nevertheless, even in this case, the increased distal latency is recognized as a key feature for the diagnosis of IgM-MGUS-related polyneuropathies [ 16 ].

In summary, the study provides valuable insights into the potential of computational algorithms for CDP classification. The discriminative features identified from the explainability analysis are consistent with established guidelines and existing literature, reinforcing the validity of our findings. This analysis suggests that more computational approaches not only validate the previously proposed criteria for electrophysiological qualitative analysis but also enhance them by providing a more nuanced and accurate discrimination among the three groups. The integration of electroneurographic variables in advanced ML models enhances our understanding of the complex nature of these disorders. When combined with clinical evaluation, these analyses hold significant potential for optimizing the diagnosis of CDP subgroups, potentially enabling earlier and more precise identification and treatment. Being the first study proposing the cross-validation of a data-driven model to discriminate CIDP and polyneuropathies associated IgM-MGUS anti-MAG +/- with > 80% accuracy, it emphasizes the need for a comprehensive investigation of ML’s potential in addressing diagnostic and prognostic problems in the spectrum of neuropathic conditions. In this promising area for research exploration to follow, our findings lay the foundation for practical clinical application, towards a data-driven refinement of neuropathy classification and then the improvement of patient care.

Limitations

Future research could aim to address some limitations identified in the study. First, although different electroneurographic variables were included, some potentially informative features were not considered (e.g., CMAP duration and temporal dispersion, both recognized markers of demyelinating processes). The addition of a more comprehensive set of electrophysiological markers may help in capturing the heterogeneity of demyelinating neuropathies. Second, the sample size may limit the statistical power of the analyses and the generalizability of the results. Although the use of nested cross-validation reduces the risk of overfitting, the inclusion of a larger cohort would improve the robustness of the proposed models. Lastly, while the study focused on neurophysiological characterization between three clinical subgroups, further analyses are needed to explore the impact of potential overlap in clinical presentation or other biomarker expression. The biological heterogeneity within subgroups may influence clinical and electrophysiological characteristics; thus, future studies involving larger cohorts enriched for additional biomarker profiles are needed to explore these aspects thoroughly. Addressing these limitations in future research will contribute to a more refined understanding of the neurophysiological profiles underlying chronic dysimmune polyneuropathies and enhance the translational value of machine learning-based diagnostic approaches.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1

Acknowledgements

NA.

Author contributions

S.B., P.L. performed analysis and developed the methodology, A.G., A.M. conceptualized the study, P.L.M., S. M. collected clinical data, A.G., M.S. performed neurophysiological examinations, S.B., P.L., drafted the initial manuscript version, A.M. provided funding for research, F.C., B.H., A.M., A.G., C.M.O. reviewed the final manuscript.

Funding

The study was supported by the Ministry of Health Current Research Funding to IRCCS Fondazione Don Carlo Gnocchi ONLUS Ricerca Corrente program” and by Regione Toscana, Bando Ricerca Salute 2018, TUNE-BEAM project (Tuscany Network for BioElectronic Approaches in Medicine).

Data availability

Data and code can be made available for research purposes upon request to the corresponding author.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethical Committee of Area Vasta Centro (CEAVC), Florence (n. 17024, 31/03/20).

Consent for publication

NA.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

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

References

1. Khadilkar SV, Deshmukh SS, Dhonde PD. Chronic dysimmune neuropathies: beyond chronic demyelinating polyradiculoneuropathy. Ann Indian Acad Neurol. 2011;14(2):81–92. 10.4103/0972-2327.82789.

2. Reynolds J, Sachs G, Stavros K. Chronic inflammatory demyelinating polyradiculoneuropathy (CIDP): clinical features, diagnosis, and current treatment strategies. R I Med J (2013). 2016;99(12):32–5.

3. Querol LA, Hartung HP, Lewis RA, et al. The role of the complement system in chronic inflammatory demyelinating polyneuropathy: implications for complement-Targeted therapies. Neurotherapeutics. 2022;19(3):864–73. 10.1007/s13311-022-01221-y.

4. Broers MC, Bunschoten C, Nieboer D, Lingsma HF, Jacobs BC. Incidence and prevalence of chronic inflammatory demyelinating polyradiculoneuropathy: A systematic review and Meta-Analysis. Neuroepidemiology. 2019;52(3–4):161–72. 10.1159/000494291.

5. Chung T, Prasad K, Lloyd TE. Peripheral neuropathy: clinical and electrophysiological considerations. Neuroimaging Clin N Am. 2014;24(1):49–65. 10.1016/j.nic.2013.03.023.

6. Ripellino P, Fleetwood T, Cantello R, Comi C. Treatment of chronic inflammatory demyelinating polyneuropathy: from molecular bases to practical considerations. Autoimmune Dis. 2014;2014:201657. 10.1155/2014/201657.

7. Talamo G, Mir MA, Pandey MK, Sivik JK, Raheja D. IgM MGUS associated with anti-MAG neuropathy: a single institution experience. Ann Hematol. 2015;94(6):1011–6. 10.1007/s00277-014-2294-7.

8. Pascual-Goñi E, Martín-Aguilar L, Lleixà C, et al. Clinical and laboratory features of anti-MAG neuropathy without monoclonal gammopathy. Sci Rep. 2019;9(1):6155. 10.1038/s41598-019-42545-8.

9. Cocito D, Durelli L, Isoardo G. Different clinical, electrophysiological and immunological features of CIDP associated with paraproteinaemia. Acta Neurol Scand. 2003;108(4):274–80. 10.1034/j.1600-0404.2003.00127.x.

10. Dalakas MC. Advances in the diagnosis, Immunopathogenesis and therapies of IgM-anti-MAG antibody-mediated neuropathies. Ther Adv Neurol Disord. 2018;11:1756285617746640. 10.1177/1756285617746640.

11. Watson JC, Dyck PJB. Peripheral Neuropathy: A Practical Approach to Diagnosis and Symptom Management. Mayo Clin Proc . 2015;90(7):940–951. 10.1016/j.mayocp.2015.05.004

12. Boukhris S, Magy L, Kabore R, et al. Atypical electrophysiologic findings in chronic inflammatory demyelinating polyneuropathy (CIDP)--diagnosis confirmed by nerve biopsy. Neurophysiol Clin. 2004;34(2):71–9. 10.1016/j.neucli.2004.01.004.

13. Motte J, Fisse AL, Köse N, et al. Treatment response to cyclophosphamide, rituximab, and bortezomib in chronic immune-mediated sensorimotor neuropathies: a retrospective cohort study. Ther Adv Neurol Disord. 2021;14:1756286421999631. 10.1177/1756286421999631.

14. Bromberg MB, Feldman EL, Albers JW. Chronic inflammatory demyelinating polyradiculoneuropathy: comparison of patients with and without an associated monoclonal gammopathy. Neurology. 1992;42(6):1157–63. 10.1212/wnl.42.6.1157.

15. Notermans NC, Franssen H, Eurelings M, Van der Graaf Y, Wokke JH. Diagnostic criteria for demyelinating polyneuropathy associated with monoclonal gammopathy. Muscle Nerve. 2000;23(1):73–9. 10.1002/(sici)1097-4598(200001)23:1%3C73::aid-mus9%3E3.0.co;2-5.

16. Joint Task Force of the EFNS and the PNS. European Federation of Neurological Societies/Peripheral Nerve Society Guideline on management of paraproteinemic demyelinating neuropathies. Report of a joint task force of the European federation of neurological societies and the peripheral nerve Society–first revision. J Peripher Nerv Syst. 2010;15(3):185–95. 10.1111/j.1529-8027.2010.00278.x.

17. Van den Bergh PYK, van Doorn PA, Hadden RDM, et al. European academy of neurology/peripheral nerve society guideline on diagnosis and treatment of chronic inflammatory demyelinating polyradiculoneuropathy: report of a joint task Force-Second revision. J Peripher Nerv Syst. 2021;26(3):242–68. 10.1111/jns.12455.

18. Broers MC, van Doorn PA, Kuitwaard K, et al. Diagnosis and treatment of chronic inflammatory demyelinating polyradiculoneuropathy in clinical practice: A survey among Dutch neurologists. J Peripher Nerv Syst. 2020;25(3):247–55. 10.1111/jns.12399.

19. Liberatore G, Manganelli F, Doneddu PE, et al. Chronic inflammatory demyelinating polyradiculoneuropathy: can a diagnosis be made in patients not fulfilling electrodiagnostic criteria? Eur J Neurol. 2021;28(2):620–9. 10.1111/ene.14545.

20. Chaudhary UJ, Rajabally YA. Underdiagnosis and diagnostic delay in chronic inflammatory demyelinating polyneuropathy. J Neurol. 2021;268(4):1366–73. 10.1007/s00415-020-10287-7.

21. Allen JA, Ney J, Lewis RA. Electrodiagnostic errors contribute to chronic inflammatory demyelinating polyneuropathy misdiagnosis. Muscle Nerve. 2018;57(4):542–9. 10.1002/mus.25997.

22. Doneddu PE, Ruiz M, Bianchi E, et al. A diagnostic score for anti-myelin-associated-glycoprotein neuropathy or chronic inflammatory demyelinating polyradiculoneuropathy in patients with anti-myelin-associated-glycoprotein antibody. Eur J Neurol. 2023;30(2):501–10. 10.1111/ene.15296.

23. Uncini A, Aretusi G, Manganelli F, et al. Electrodiagnostic accuracy in polyneuropathies: supervised learning algorithms as a tool for practitioners. Neurol Sci. 2020;41(12):3719–27. 10.1007/s10072-020-04499-y.

24. Haque F, Bin Ibne Reaz M, Chowdhury MEH, et al. Performance analysis of conventional machine learning algorithms for diabetic sensorimotor polyneuropathy severity classification. Diagnostics (Basel). 2021;11(5):801. 10.3390/diagnostics11050801.

25. Alarcón-Narváez D, Hernández-Torruco J, Hernández-Ocaña B, Chávez-Bosquez O, Marchi J, Méndez-Castillo JJ. Toward a machine learning model for a primary diagnosis of Guillain-Barré syndrome subtypes. Health Inf J. 2021;27(3):14604582211021471. 10.1177/14604582211021471.

26. Chang CW, Ro LS, Lyu RK, et al. Establishment of a new classification system for chronic inflammatory demyelinating polyneuropathy based on unsupervised machine learning. Muscle Nerve. 2022;66(5):603–11. 10.1002/mus.27702.

27. Baek SH, Hong YH, Choi SJ, et al. Electrodiagnostic data-driven clustering identifies a prognostically different subgroup of patients with chronic inflammatory demyelinating polyneuropathy. J Neurol Neurosurg Psychiatry. 2019;90(6):674–80. 10.1136/jnnp-2018-319758.

28. Scarpino M, Lolli F, Carrai R, et al. Diagnostic accuracy of neurophysiological criteria for early diagnosis of AIDP: A prospective study. Neurophysiol Clin. 2016;46(1):35–42. 10.1016/j.neucli.2015.12.008.

29. Pedregosa F, Varoquaux G, Gramfort A et al. Scikit-learn: machine learning in Python. J Mach Learn Res. 2011;12:2825–2830. http://jmlr.org/papers/v12/pedregosa11a.html

30. Akiba T, Sano S, Yanase T, Ohta T, Koyama M, Optuna. A Next-generation hyperparameter optimization framework. Published Online July. 2019;25. 10.48550/arXiv.1907.10902.

31. Zou H, Hastie T. Regularization and variable selection via the elastic net. J Royal Stat Soc Ser B: Stat Methodol. 2005;67(2):301–20. 10.1111/j.1467-9868.2005.00503.x.

32. Cortes C, Vapnik V. Support-vector networks. Mach Learn. 1995;20(3):273–97. 10.1007/BF00994018.

33. Cover T, Hart P. Nearest neighbor pattern classification. IEEE Trans Inf Theory. 1967;13(1):21–7. 10.1109/TIT.1967.1053964.

34. Breiman L. Random forests. Mach Learn. 2001;45(1):5–32. 10.1023/A:1010933404324.

35. Breiman L, Friedman J, Olshen RA, Stone CJ. Classification and regression trees. Chapman and Hall/CRC; 2017. 10.1201/9781315139470.

36. Lundberg SM, Erion G, Chen H, et al. From local explanations to global Understanding with explainable AI for trees. Nat Mach Intell. 2020;2(1):56–67. 10.1038/s42256-019-0138-9.

37. Khwaja J, D’Sa S, Minnema MC, Kersten MJ, Wechalekar A, Vos JM. IgM monoclonal gammopathies of clinical significance: diagnosis and management. Haematologica. 2022;107(9):2037–50. 10.3324/haematol.2022.280953.

38. Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med. 2022;28(1):31–8. 10.1038/s41591-021-01614-0.

39. Isoardo G, Migliaretti G, Ciaramitaro P, et al. Differential diagnosis of chronic dysimmune demyelinating polyneuropathies with and without anti-MAG antibodies. Muscle Nerve. 2005;31(1):52–8. 10.1002/mus.20230.

40. Shapley LS. 17. A value for n-Person games. 17. A value for n-Person games. Princeton University Press; 2016. pp. 307–18. 10.1515/9781400881970-018.

41. Campagnini S, Liuzzi P, Mannini A, et al. Cross-validation of predictive models for functional recovery after post-stroke rehabilitation. J Neuroeng Rehabil. 2022;19(1):96. 10.1186/s12984-022-01075-7.

42. Towman FH, Hamdan FB. Electrodiagnostic patterns of demyelination and Hughes functional grading in typical chronic inflammatory demyelinating polyneuropathy. Neurol Sci Neurophysiol. 2020;37(2):50–6. 10.4103/NSN.NSN_8_20.

43. Bromberg MB. Review of the evolution of electrodiagnostic criteria for chronic inflammatory demyelinating polyradicoloneuropathy. Muscle Nerve. 2011;43(6):780–94. 10.1002/mus.22038.

44. Rajabally YA, Varanasi S. Practical electrodiagnostic value of F-wave studies in chronic inflammatory demyelinating polyneuropathy. Clin Neurophysiol. 2013;124(1):171–5. 10.1016/j.clinph.2012.05.019.

45. Koski CL, Baumgarten M, Magder LS, et al. Derivation and validation of diagnostic criteria for chronic inflammatory demyelinating polyneuropathy. J Neurol Sci. 2009;277(1–2):1–8. 10.1016/j.jns.2008.11.015.

46. Rajabally YA, Nicolas G, Piéret F, Bouche P, Van den Bergh PYK. Validity of diagnostic criteria for chronic inflammatory demyelinating polyneuropathy: a multicentre European study. J Neurol Neurosurg Psychiatry. 2009;80(12):1364–8. 10.1136/jnnp.2009.179358.

Associated Data

Supplementary Materials

Supplementary Material 1

Data Availability Statement

Data and code can be made available for research purposes upon request to the corresponding author.