Article Text

Download PDFPDF

Research paper
Electrodiagnostic data-driven clustering identifies a prognostically different subgroup of patients with chronic inflammatory demyelinating polyneuropathy
  1. Seol-Hee Baek1,
  2. Yoon-Ho Hong2,
  3. Seok-Jin Choi3,
  4. So Hyun Ahn4,
  5. Kee Hong Park5,
  6. Je-Young Shin4,
  7. Jung-Joon Sung4
  1. 1 Department of Neurology, Korea University Medical Center, Korea University College of Medicine, Seoul, Republic of Korea
  2. 2 Department of Neurology, Seoul Metropolitan Government Boramae Medical Center, Seoul National University College of Medicine, Neuroscience Research Institute, Seoul National University Medical Research Council, Seoul, Republic of Korea
  3. 3 Department of Neurology, Inha University Hospital, Incheon, Republic of Korea
  4. 4 Department of Neurology, Neuroscience Research Institute, Seoul National University Hospital, Seoul National University College of Medicine, Seoul, Republic of Korea
  5. 5 Department of Neurology, Gyeongsang National University Hospital, Jinju, Republic of Korea
  1. Correspondence to Professor Jung-Joon Sung, Department of Neurology, Seoul National University Hospital, Seoul National University College of Medicine, Seoul 03080, Republic of Korea; jjsaint66{at}


Objective This study aimed to explore the correlations between electrodiagnostic (EDX) features in patients with chronic inflammatory demyelinating polyneuropathy (CIDP) and to investigate whether EDX data-driven clustering can identify a distinct subgroup regarding clinical phenotype and treatment response.

Methods We reviewed clinical and EDX data of 56 patients with definite CIDP fulfilling the 2010 European Federation of Neurological Societies and Peripheral Nerve Society criteria at two teaching hospitals. A hierarchical agglomerative clustering algorithm with complete linkage was used to partition the patients into subgroups with similar EDX features. A stepwise logistic regression analysis was performed to evaluate predictors of the long-term outcome.

Results EDX data-driven clustering partitioned the patients into two clusters, identifying a distinct subgroup characterised by coexistence of prominent conduction slowing and markedly reduced distally evoked compound muscle action potential (CMAP) amplitudes. This cluster of patients was significantly over-represented by an atypical subtype (distal acquired demyelinating symmetric polyneuropathy) compared with the other cluster (70% vs 26.1%, p=0.042). Furthermore, patients in this cluster invariably showed favourable long-term treatment outcome (100% vs 63%, p=0.023). In logistic regression analyses, the initial disability (OR 6.1, 95% CI 2.4 to 25.4), F-wave latency (OR 0.93, 95% CI 0.86 to 0.98) and distal CMAP duration (OR 0.96, 95% CI 0.91 to 0.99) were significant predictors of the poor long-term outcome.

Conclusion Our results show that EDX data-driven clustering could differentiate a pattern of EDX features with prognostic implication in patients with CIDP. Reduced distally evoked CMAPs may not necessarily predict poor responses to treatment, and active treatment is warranted when prominent slowing of conduction is accompanied in the distal segments.

  • chronic inflammatory demyelinating polyneuropathy
  • electrodiagnosis
  • clustering
  • treatment
View Full Text

Statistics from


Chronic inflammatory demyelinating polyneuropathy (CIDP) is the most common autoimmune chronic polyneuropathy with an estimated prevalence of approximately 5 per 100 000 persons according to current diagnostic criteria.1 CIDP is usually treated effectively with corticosteroids, intravenous immunoglobulin (IVIg) and plasma exchange, but 10% to 30% of patients have the difficult-to-control disease and eventually have severe disability.2–4 Several clinical and laboratory features, including initial disability at diagnosis, age at onset, gender, disease duration, clinical subtype, muscle atrophy, cerebrospinal fluid (CSF) protein and the presence of autoantibodies, reportedly affect treatment response and long-term disability.5–14

Electrodiagnostic (EDX) studies can provide important diagnostic measures of peripheral nerve demyelinating pathology. The current EDX criteria recommended by the European Federation of Neurological Societies and Peripheral Nerve Society (EFNS/PNS) require demonstration of demyelinating abnormalities for the diagnosis of CIDP.15 Regarding the prognostic value of EDX studies, some EDX features have been reported to predict treatment response in patients with CIDP.7 11 16–18 Decreased compound muscle action potentials (CMAPs) were reported to be pronounced in non-responders of IVIg.7 11 16 A higher number and degree of demyelinating features have been demonstrated to predict higher treatment response rates.17 18 The distribution patterns of demyelination reportedly correlate with clinical profiles and treatment response as well.10 19 However, previous studies have analysed individual EDX parameters separately; therefore, these studies are limited in capturing complex patterns within a metric of EDX conduction parameters. The present study investigated the relationships between conduction parameters in patients with CIDP, and used an unsupervised clustering approach to examine whether EDX data-driven clustering can identify a distinct subgroup of patients with regards to clinical phenotypes and response to treatment.



We reviewed medical records of adult patients who were diagnosed with CIDP at two teaching hospitals (Seoul National University Hospital and Seoul Metropolitan Government Boramae Medical Center) between January 2004 and August 2015. Patients were excluded from this study if their final diagnosis turned out to be not CIDP, and the non-CIDP groups included hereditary demyelinating neuropathy (n=12), anti-myelin-associated glycoprotein neuropathy (n=3), multifocal motor neuropathy (n=5), polyneuropathy, organomegaly, endocrinopathy, monoclonal protein, skin changes syndrome (n=5), amyloid neuropathy (n=3) and lymphoma (n=1). To ensure diagnostic certainty, this study included only patients with definite CIDP, who fulfilled the diagnostic criteria proposed by the Joint Task Force of the EFNS/PNS, and excluded probable and possible patients with CIDP (n=4 and n=1, respectively). Ten patients were also excluded because of unavailable EDX data at the time of diagnosis.

Clinical assessment

The collected clinical and laboratory data included demographics, neurological manifestations at diagnosis, time from symptom onset to diagnosis, follow-up duration, disability as measured by modified Rankin disability scale (mRS), comorbidities, treatments, CSF protein and monoclonal gammopathy. Typical and atypical subtypes of CIDP were determined by the EFNS/PNS clinical diagnostic criteria.15 Atypical CIDP included multifocal acquired demyelinating sensory and motor neuropathy (MADSAM), distal acquired demyelinating symmetric polyneuropathy (DADS) and pure sensory CIDP. Because no formal clinical diagnostic criteria were available for these atypical subtypes, we adopted the following operational definitions as described in previous studies. MADSAM is defined as a typical mononeuropathy multiplex, and the asymmetry of weakness is determined as differences in muscle strength by one or more Medical Research Council (MRC) scales in the homonymous muscles.10 DADS is defined as a predominantly distal and symmetric sensorimotor polyneuropathy sparing proximal limb muscles.9 Patients were assigned to pure sensory CIDP if they presented with purely sensory symptoms and signs regardless of abnormalities in motor nerve conduction studies.20 Treatment outcomes were determined based on the disability measured at the latest visit following treatments. The outcomes are considered favourable when the mRS score is 2 (slight disability; unable to carry out all previous activities, but able to look after own affairs without assistance) or less, and otherwise unfavourable.

EDX tests

A nerve conduction study (NCS) was performed using the standard techniques of percutaneous supramaximal stimulation and surface electrode recording with standard electromyography (EMG) equipment. Limb skin temperature was maintained above 31°C. Motor NCS was performed in the median, ulnar, peroneal and posterior tibial nerves, and the measured conduction parameters included CMAP negative peak amplitude, distal motor latency (DML), distal CMAP negative peak duration (DUR), nerve conduction velocity (NCV), conduction block (CB), temporal dispersion (TD) and minimal F-wave latency (FL). The distal stimulation distance was 5 cm for the median and ulnar nerves, 8 cm for the peroneal nerve and 10 cm for the posterior tibial nerve. Motor NCV was assessed in the segments of wrist to elbow for the median nerve, wrist to elbow and across the elbow for the ulnar nerve, ankle to fibular head and across the fibular head for the peroneal nerve, and ankle to popliteal fossa for the posterior tibial nerve. FL was corrected for the patient’s age and height.21 CB was defined as amplitude reduction (%) of the proximal negative peak CMAP relative to distal, and TD as duration increase (%) between the proximal and negative peak CMAP. Sensory nerve conduction was measured orthodromically in the median and ulnar nerves and antidromically in the sural nerve. High-frequency and low-frequency filters were set at 10 kHz and 20 Hz, respectively. All patients underwent the EDX evaluation at three or four limbs. EDX data were analysed if the patients fulfil the 2010 EFNS/PNS criteria for CIDP.15

Hierarchical clustering

EDX raw data were converted to percentage values relative to upper or lower limits of normal (reference values in our EMG laboratory are provided in online supplementary file). Then, the mean value of each conduction parameter was calculated across all tested motor nerves for explorative correlation analysis and clustering. Missing data were imputed by a random forest algorithm using the missForest R package.22

Supplementary data

Cluster analysis, an unsupervised machine learning approach, is a collection of techniques for finding subgroups or clusters in a data set. The groupings are constructed such that the observations within each group are quite similar to each other, while observations in different groups are quite different from each other. We used an agglomerative hierarchical clustering approach with EDX data in order to partition the patients into distinct subgroups based on the similarity of EDX features.23 In this approach, each patient is treated as its own cluster in the beginning, and pairs of clusters are combined sequentially into larger clusters based on a measure of dissimilarity between the clusters. Euclidean distance of the normalised EDX data was used for dissimilarity measure, and the maximal intercluster distance was used to define the distance between clusters (complete linkage). Determining the optimal number of clusters is inherently a challenging task in cluster analysis. A number of validity indices have been proposed to evaluate the internal validity of clusters and also to determine the relevant number of clusters. Unfortunately, however, there is no consensus on a single best approach. In this study, the interval validity, which refers to how well the clustering results fit the data set, and hence the optimal number of clusters was determined using the NbClust R package in which a variety of validity indices (30 in total) are implemented.24 Each validity index proposes the optimal number of clusters, based on the information about compactness within clusters and separation between clusters, as well as other factors, such as geometric or statistical properties of the data.24 The best number of clusters was determined according to the majority rule, in which the number of clusters is proposed as the optimal clustering scheme by the most validity indices. On the other hand, the external validity was evaluated by comparing the results of cluster analysis to externally provided data including clinical phenotypes, laboratory features and treatment outcome.

Statistical analysis

Descriptive summaries are presented as frequency and proportion for categorical variables and median and range for continuous variables. Correlations between conduction parameters were evaluated by Pearson correlation coefficients. Statistical comparisons were performed using χ2 or Fisher’s exact test for categorical variables and Mann-Whitney U test for continuous variables. A multivariate logistic regression analysis was performed to identify potential predictors of the long-term treatment outcomes (favourable vs unfavourable). Clinical and EDX variables with p value of <0.1 were selected by the univariate logistic regression analysis, and the selected variables were included in the backward stepwise multivariate logistic regression analysis. A p value of <0.05 was considered statistically significant. All statistical analyses were performed using R V.


Clinical profiles of patients

A total of 56 patients (women, 37.5%) with definite CIDP were included in this study. Clinical profiles of the patients are summarised in table 1. Median age at diagnosis was 54 years (range, 20–84 years). Median time from symptom onset to diagnosis was 7 months. These 56 cases of CIDP were reclassified as 22 typical CIDP (39%), 19 DADS (34%), 10 MADSAM (18%) and 5 pure sensory CIDP (9%). Initial disability at diagnosis was mild to moderate (MRS score, 2; range, 1–5). All patients were initially treated with high-dose corticosteroids (intravenous methylprednisolone, 1000 mg/day for 3 to 5 days) or IVIg (0.4 g/kg/day for 5 days). Immunosuppressive agents other than corticosteroids were administered to 19 patients (34%), and these agents included azathioprine, mycophenolate mofetil, tacrolimus, cyclophosphamide and rituximab. Thirty-nine patients (70%) achieved favourable outcomes (mRS, 2 or less) during the median follow-up of 44 months (IQR, 18.5–89.8).

Table 1

Comparisons of clinical profiles between two clusters of patients partitioned by electrodiagnostic data-driven clustering

Exploratory analysis of EDX data

We investigated the correlations between EDX conduction parameters (figure 1). A positive correlation was demonstrated between DML and DUR (demyelinating features in the distal segments; r=0.53, p<0.001). As for the demyelinating features in the intermediate segments, a positive correlation was found between CB and TD (r=0.26, p=0.049), and negative correlations of NCV with CB (r=−0.32, p=0.017) and TD (r=−0.29, p=0.033). FL was found to correlate positively with DML (r=0.41, p=0.0018) and DUR (r=0.29, p=0.028) and negatively with NCV (r=−0.60, p<0.001) but not with other parameters. Distal CMAP amplitude, which is traditionally regarded as an indicator of axonal damage, was negatively correlated with DML (r=−0.45, p<0.001) and positively with NCV (r=0.38, p=0.004), but not with other demyelinating parameters, including DUR, CB, TD and FL. A correlation network graph was constructed to facilitate understanding of a whole picture of these complex relationships among EDX conduction parameters (online supplementary figure 1).

Supplementary data

Figure 1

Multiple scatter plots and correlations between electrodiagnostic conduction parameters. Values of each conduction parameter are expressed as percentages of upper or lower limits of normal except for conduction block (CB) and temporal dispersion (TD). Each panel on the lower triangle is a scatter plot with smooth curve fitted by Loess (locally estimated scatterplot smoothing) for a pair of variables whose identities are given by the corresponding row and column labels. Pearson correlation coefficient for each pair was presented on the corresponding panel on the upper triangle. ***p<0.001; **p<0.01; *p<0.05. CMAP, compound motor action potential; DML, distal motor latency; DUR, distal CMAP duration; FL, F-wave latency; NCV, nerve conduction velocity.

EDX data-driven clustering

Using hierarchical clustering, we sought to identify distinct subgroups of patients who were similar in the EDX pattern. The prior unknown optimal number of clusters was determined based on the results of various internal validity measures (online supplementary table 1). There was no unanimous choice regarding the optimal number of clusters. Among 26 validity indices, 7 indices proposed 2 as the best number of clusters, 5 indices proposed 4, and 4 indices proposed 7, and so on. We determined the optimal number of clusters according to the majority rule, partitioning the patients into two clusters. Figure 2 shows the results of clustering analysis and heat map of EDX data.

Figure 2

Hierarchical clustering and heat map display of electrodiagnostic (EDX) data obtained from 56 patients with definite chronic inflammatory demyelinating polyneuropathy. Patient group memberships (cluster and clinical subtype) are indicated by row annotations of different colours above the heat map and below the dendrogram. Patient clustering was performed on the normalised mean data of 7 EDX parameters using Euclidean distance and complete linkage. CB, conduction block; CMAP, compound motor action potential amplitude; DADS, distal acquired demyelinating symmetric polyneuropathy; DML, distal motor latency; DUR, distal CMAP duration; FL, F-wave latency; MADSAM, multifocal acquired demyelinating sensory and motor neuropathy; NCV, nerve conduction velocity; TD, temporal dispersion.

Comparisons between clusters

Comparison between two clusters was performed for each EDX parameter (figure 3, online supplementary table 2). Compared with cluster 1 (n=46), cluster 2 (n=10) had more prominent demyelinating features with significantly prolonged DML, DUR and FL and reduced NCV (p<0.01). Notably, the distal CMAP amplitude was significantly lower in cluster 2 than in cluster 1 (p<0.01). CB and TD were not significantly different between the two clusters.

Figure 3

Comparisons of conduction parameters between two electrodiagnostic data-driven clusters. Values are expressed as percentages of upper or lower limit of normal. CB, conduction block; CMAP, compound motor action potential amplitude; DML, distal motor latency; DUR, distal CMAP duration; FL, F-wave latency; NCV, nerve conduction velocity; TD, temporal dispersion.

Subsequently, we analysed the clinical profiles of two EDX data-driven clusters (table 1). No significant differences were found in gender, age, time from onset to diagnosis, follow-up duration, comorbid diabetes mellitus, major clinical and laboratory features, treatments and initial disability. However, a significant difference was noted in the clinical subtypes (p=0.042). In cluster 1, the most common subtype was typical CIDP (41.3%), followed by DADS (26.1%), MADSAM (21.7%) and pure sensory CIDP (10.9%). In contrast, in cluster 2, the most common subtype was DADS (70.0%), followed by typical CIDP (30.0%). No patients with MADSAM or pure sensory CIDP were observed in cluster 2. In addition, the long-term treatment outcomes were significantly different between these two clusters (p=0.023). After the median follow-up of 44 months, 17 patients (37%) in cluster 1 were in moderate disability or worse (mRS ≥3), but all patients in cluster 2 had favourable outcomes (mRS ≤2).

Long-term outcome predictors

We then performed logistic regression analyses to evaluate potential predictors of the long-term treatment outcomes. Univariate logistic regression tests revealed possible associations of the long-term treatment outcome with initial mRS, MRC sum score at diagnosis, IVIg treatment and EDX parameters including DML, DUR, NCV and FL (p≤0.1; table 2). A backward stepwise regression analysis with these variables revealed initial mRS (OR 6.1), FL (OR 0.93) and DUR (OR 0.96) as significant predictors of the long-term treatment outcomes (table 2).

Table 2

Univariate and stepwise backward multivariate logistic regression analyses


The present study showed that EDX data-driven cluster analysis could partition the patients with CIDP into subgroups with a distinct pattern of EDX features. Notably, clinical subtypes and long-term treatment outcomes were significantly different between the two EDX data-based patient clusters, indicating the phenotypic and prognostic implications of the unsupervised approach.

Several studies have found the association between EDX features and treatment outcomes in patients with CIDP. Reduced distally evoked CMAPs, traditionally regarded as a marker for secondary axonal damage in CIDP, have been demonstrated to predict the lack of response to IVIg treatment.7 On the contrary, demyelinating features have been reported to be associated with high treatment response rates.7 10 18 19 Conduction block is found to be more frequent in treatment responders.7 Fulfilling a high number of EDX demyelinating criteria is associated with high treatment response rates, and high degrees of demyelination are associated with treatment responsiveness in patients with comorbid diabetes mellitus.18 The distribution pattern of demyelination along the course of the nerve has been recently claimed to have a prognostic implication, and demyelination predominant in the distal nerve segments is associated with a good response to treatment.10 19

Our results are in line with those in previous studies, in which prominent demyelinating features, particularly in the distal segments but not limited to those, are associated with favourable outcomes. However, the present study suggests that reduced distally evoked CMAP may not necessarily represent secondary axonal damages. Indeed, distal CMAP amplitudes were significantly lower in the subgroup of patients with more prominent demyelinating features. Considering the favourable treatment outcomes in this subgroup, it would be plausible that conduction block and/or temporal dispersion in the distal segments might cause the reduced distal CMAPs. Dyck et al demonstrated significant increases in distal CMAP amplitudes after successful treatment with plasmapheresis or IVIg in patients with CIDP.26 Resolution of conduction block in the distal nerve segments could account for the rapid large improvement. We propose that active treatment should be considered even in those patients with reduced distally evoked CMAPs, particularly when markedly demyelinating features are accompanied in the distal segments, such as prolongation of DML and DUR. Needle EMG would also help to differentiate secondary axonal damage and demyelinating conduction block in patients with reduced distal CMAPs. Further studies are needed to investigate the temporal changes of EDX parameters and to elucidate the pathological correlates of reduced distal CMAPs in patients with CIDP patients.

The present study suggests that initial disability is a strong predictor of long-term treatment outcomes, and the results are in line with those in previous studies.5 6 Notably, initial disability was not significantly different between the two EDX-based patient clusters, indicating that EDX clustering is not related to the initial disability. Among EDX parameters, only FL and DUR were significantly associated with treatment outcomes, but the results should be interpreted with caution because of the marginal ORs. Inconsistency has been observed in the associations between treatment outcomes and EDX parameters6 27 and could be explained, in part at least, by the complex relationship among various EDX parameters. Multicollinearity between conduction parameters in CIDP could undermine the robustness of any regression models. It should also be pointed out that the significant association between cluster membership and treatment outcome was revealed in Fisher’s exact test, but not in univariate logistic regression analysis. The discrepancy may be explained by the difference of the two tests concerning the statistical power particularly for a small sample size.

Another interesting finding in this study is the association between EDX data-based patient clusters and clinical subtypes. Of note, the cluster, which shared a distinct EDX pattern of prominent distal conduction slowing and markedly reduced distal CMAP, was significantly over-represented by DADS compared with the other cluster. Neither MADSAM nor pure sensory CIDP was found in this cluster. Distribution patterns of conduction abnormalities was demonstrated to determine clinical phenotypes.19 It was also reported that electrophysiological profiles and treatment response were different between typical CIDP and MADSAM.10 Our results corroborate and expand findings of previous reports on the clinical–electrophysiological correlations in typical and atypical CIDP. Given the potential of cluster analysis to redefine patient subgroups based on EDX profiles, it would be interesting to investigate whether EDX data-driven clustering can identify distinct subgroups within a more homogeneous patient population such as typical CIDP. Further studies with a large number of patients are needed to investigate the clinical–electrophysiological correlations at a more granular level.

The proportion of typical CIDP varied widely in previous studies, ranging from 46% to 80.4%.1 4 10 28 29 Atypical subtypes including DADS, MADSAM and pure sensory CIDP also varied largely in proportion (5%–48.3%, 8%–34% and 1%–23.9%, respectively).1 4 10 28 29 Compared with previous studies, the proportions of typical CIDP and DADS in our cohort seemed to be relatively low (39.3%) and high (33.9%), respectively. Selection bias might account for the difference, although our cohort consisted of consecutive patients. The variation across studies might also be explained by the lack of established clinical criteria for atypical CIDP. DADS is defined as predominantly distal, but the exact proximal to distal gradient of motor deficit was not specified in the 2010 EFNS/PNS criteria. More specific consensus criteria for atypical CIDP may help to identify the sources of variation concerning the frequency of clinical subtypes in the literatures.

This study is retrospective and uncontrolled. Other limitations include the use of modified Rankin scale that does not provide information on arm function and hence would not be the optimal disability measure in immune-mediated polyneuropathy such as CIDP. It should be also acknowledged that the use of unadjusted p values in multiple comparisons between clusters might entail the risk of type I error inflation, and that our results need to be confirmed in larger studies. Lastly, we aggregated multiple values from different nerves to form a single mean value for each conduction parameter. This might lead to information loss, particularly on the nerve predilection in cases of MADSAM. However, our approach enables the clustering based on the distribution patterns of conduction abnormalities along the course of the nerves, circumventing the problem of high dimensional data in cluster analysis. The problem, often referred to as the ‘curse of dimensionality’, concerns the effect of dimensionality on the distance between observations; the distance between observations tends to become relatively uniform as dimensionality increases.30

On the other hand, the strength of this study is that it included exclusively the patients with definite CIDP fulfilling the 2010 EFNS/PNS criteria to ensure diagnostic certainty. Treatment can affect EDX features, so we also restricted EDX analysis to those data obtained at the time of diagnosis and before initial treatment. Our results confirm the multiple correlations between EDX conduction parameters in patients with CIDP and raise the need for machine learning–based pattern recognition approaches. In this study, hierarchical clustering of EDX data was used to partition the patients into two subgroups with similar EDX pattern, and a distinct subgroup with favourable treatment outcomes was identified. The present results suggest that reduced distally evoked CMAPs may not necessarily indicate poor responses to treatment but represent demyelinating pathology in the distal segments, which could resolve with active treatment.


View Abstract


  • S-HB and Y-HH contributed equally.

  • Contributors S-HB: acquisition and statistical analysis of data and drafting of the manuscript. Y-HH: study concept and design, clustering analysis, statistical analysis of data and drafting of the manuscript. S-JC: acquisition of data and clustering analysis. SHA, KHP and J-YS: acquisition of data. J-JS: study concept and design and critical revision of the manuscript.

  • Funding This study was supported by grants from the Korea Healthcare Technology R&D project, Ministry of Health and Welfare, Republic of Korea (HI14C3347) and the National Research Foundation of Korea (NRF) grant funded by the Korean Government (MSIP) (NRF-2018R1A5A2025964).

  • Competing interests None declared.

  • Patient consent for publication Not required.

  • Ethics approval This study was approved by the institutional review boards of Seoul National University Hospital (IRB H-1603-096-750) and Seoul Metropolitan Government Boramae Medical Center (IRB 16-2016-57).

  • Provenance and peer review Not commissioned; externally peer reviewed.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.