Objective characterization of hip pain levels during walking by combining quantitative electroencephalography with machine learning

Abstract

Pain is an undesirable sensory experience that can induce depression and limit individuals' activities of daily living, in turn negatively impacting the labor force. Affected people frequently feel pain during activity; however, pain is subjective and difficult to judge objectively, particularly during activity. Here, we developed a system to objectively judge pain levels in walking subjects by recording their quantitative electroencephalography (qEEG) and analyzing data by machine learning. To do so, we enrolled 23 patients who had undergone total hip replacement for pain, and recorded their qEEG during a five-minute walk via a wearable device with a single electrode placed over the Fp1 region, based on the 10-20 Electrode Placement System, before and three months after surgery. We also assessed subject hip pain using a numerical rating scale. Brain wave amplitude differed significantly among subjects with different levels of hip pain at frequencies ranging from 1 to 35 Hz. qEEG data were also analyzed by a support vector machine using the Radial Basis Functional Kernel, a function used in machine learning. That approach showed that an individual's hip pain during walking can be recognized and subdivided into pain quartiles with 79.6% recognition Accuracy. Overall, we have devised an objective and non-invasive tool to monitor an individual's pain during walking.

Conflict of interest statement

The authors declare no competing interests.

Figures

Figure 1

Development of an EEG recording system applicable to walking. (a) The wearable EEG recording device, MindWave Mobile II (Neurosky), used in this study to record the EEG from electrode position Fp1. Conductive material reference and ground electrodes were contained within the ear clip. (b)Wireless control of the device with an iPad. Subjects were able to walk with hands free. (c) Task diagram. The task consisted of a one-minute rest block with subjects in a sitting position with eyes closed, then a walking block of 5 min, and finally a rest block like the first. For the walking block, we encouraged patients to walk at a uniform pace and in a way that simulated their daily habits. (d) Pain classification flowchart. We asked about hip pain while recording EEG, before and after THA surgery. We collected 46 sets of NRS data from 23 subjects, and sets were subdivided into Pain (–) (NRS 0, n = 16) or Pain ( +) (NRS ≥ 1, n = 30) groups. Pain ( +) groups were subdivided into quartiles, as indicated. Each EEG data recording session of 300 s from 46 individuals was divided into 10 data sets of 30 s. Thus the total number of datasets analyzed was 460.

Figure 2

Flow chart of study subjects. Eight patients were excluded due to refusal to undergo follow-up examination, three were excluded by non-standard course such as re-operation, and six were excluded due to incomplete data sets. Most patients who refused undergo an examination were unable to walk continuously for 5 min.

Figure 3

Brain wave amplitude at various frequencies differs significantly between Pain (−) and (+). Brain wave amplitude was recorded during a five-minute walk in patients with (Pain (+), NRS > 1, n = 30) or without (Pain (−), NRS 0, n = 16) hip pain, and the difference in amplitude at frequencies ranging from 1 to 35 Hz was compared between groups. Brain wave amplitude was significantly higher in some frequencies in the Pain (+) than Pain (−) groups (*p < 0.05).

Figure 4

Brain wave amplitude differs significantly among pain levels. Brain wave amplitude was recorded during a five-minute walk in patients with or without hip pain. Each pain level was subdivided into the following quartiles based on the NRS: none (NRS 0, n = 16), mild (NRS 1–3, n = 8), moderate (NRS 4–6, n = 15) and severe (NRS 7–10, n = 7), and the power spectrum at frequencies ranging from 1 to 35 Hz was analyzed.

Figure 5

Kellgren/Lawrence grades are associated with NRS before THA surgery. Hip pain, as determined by NRS, was significantly greater in patients with K/L = 4 than those with K/L = 3. One osteonecrosis patient was excluded from analysis because their K/L grade was a classification adapted to OA, and osteonecrosis is known to cause strong pain with less radiographic change.