Low back pain precedes the development of new knee pain in the elderly population; a novel predictive score from a longitudinal cohort study

The aims of this study were to investigate the association between knee and low back pain/disabilities and to develop a predictive score that enables the identification of those who are likely to develop new knee pain within a period of 5 years. We selected participants aged ≥ 50 years because it has been shown that the prevalence of knee OA increases dramatically over the age of 50 [1, 2].

This prospective, longitudinal study included part of the general, comprehensive cohort that was recruited from the general population living in Nagahama, a largely rural city of 125,000 inhabitants in Shiga Prefecture, located in central Japan, and which has been reported elsewhere [10, 11]. We recruited residents for this particular study between 2007 and 2010. The inclusion criteria for the study were as follows: (1) aged ≥ 50 years at the time of the first survey, (2) able to participate independently in the health examination, (3) having no difficulties in communication, and (4) voluntarily deciding to participate in the project. This study was designed in accordance with the Helsinki Declaration and was approved by the Ethics Committee of Kyoto University Graduate School and Faculty of Medicine (No. C278). Written informed consent for this study was obtained from all participants.

A total of 5932 people aged ≥ 50 years agreed to participate in this study at the first surveillance from 2007 to 2010. Then, the second survey was sent in 2015 to all respondents of the first survey, and 5576 participants returned the form (94.0% response). The paired data from the two time points were analyzed, and the 5046 subjects whose total pain score of the Japanese Knee Osteoarthritis Measure (JKOM) [12] increased by 3 or more (9% net increase in a possible 32 points) were analyzed as a “new symptom” group, based on a previous similar report [13].

The presence of knee pain was determined by a patient-reported outcome score, JKOM, which was established and validated previously [12]. The pain score consists of eight subscales, in each of which, a subject chooses no to severe symptom. No symptom is regarded as score 0, and severe symptom is score 4. Subjects whose scores were 0 or 1 in all the eight pain subscales were included in subsequent analyses as “no symptom” subjects.

Basic clinical parameters were measured and surveyed at baseline. Blood and urine samples were also collected. Age at the time of the first survey, sex, and BMI based on height and weight at the first survey were recorded. Information about clinical history, smoking, and drinking habits was obtained using a structured questionnaire. The weight change between the time of the first survey and when the participant was 20 years old was reported by the participant in five categories as no change (< 3 kg), slight increase (3–10 kg), substantial increase (> 10 kg), slight decrease (3–10 kg), and substantial decrease (> 10 kg). Participants were asked whether they were never smokers, had stopped smoking, or were current smokers. Current smokers were asked how many cigarettes they smoked per day and how many years they had smoked. Participants were also asked whether they were never regular drinkers, had stopped drinking, or drank currently. Current drinkers were asked how much they consumed per day using self-calculation of total units, with a unit equaling a bottle of beer (500 ml), a glass of wine (240 ml), or a shot of liquor (180 ml) [14]. Participants were asked to classify how much they moved in daily life as sedentary (mostly sitting), moderately active (sometimes did walking, shopping, or light sport) or active (played sports or did exercise regularly). The severity of low back pain and its disabilities was evaluated by the Roland–Morris disability questionnaire (RMDQ) [15]. Mental health was surveyed using a subscale of the Short-Form 36-Item Health Survey (SF36) [16]. We also measured the ankle-brachial pressure index (ABI) as a marker of vessel aging, serum high-sensitivity C-reactive protein (hsCRP) as a marker of inflammation, and urine cross-linked N-telopeptide of type I collagen (NTX) and urine C-terminal telopeptide of type I collagen (CTX) as markers of osteoporosis.

To conduct multivariate analyses, we divided the entire group into the two for mental health and low back pain by the median score, respectively. We performed logistic regression analysis to calculate the regression coefficients of the predictor variables. The backwards stepwise selection method was used to reduce the number of predictors incorporated into the final model, the aim being model simplicity. The significance level for removing variables from the model was set at P ≥ 0.20. We employed a scoring system to present the final model. Each predictor regression coefficient was divided by twice the smallest regression coefficient and rounded to the nearest integer. We calculated the risk of pain worsening at 5 years as e^lp/(1 + e^lp), where lp is the linear predictor for each subject. Prediction model made by regression analysis is known to overfit the data and have problem in terms of generalizability. Therefore, we applied a shrinkage factor to the regression coefficients when calculating risk of knee pain worsening for better prediction (Additional file 1: Supplementary note) [17, 18]. To measure the performance of the model, we used the C-index to assess its discriminatory ability. We also evaluated the calibration by plotting the predicted risk in deciles against the corresponding proportion of the subjects who experienced worsening of knee pain.

Figure 1 shows the process for selection of the participants. As planned, we selected the 5046 respondents with negligible knee pain (90.5%). The data for participants whose follow-up period was < 4.5 or > 6.5 years were excluded from further analyses (408, 8.1%), resulting in 4638 participants being analyzed. The average follow-up period was 5.4 years (4.8–6.2 years).

At baseline, the average age of the participants was 62.4 years, and the majority were women (64.8%). Their average BMI was 22.5, which was comparable to that reported elsewhere in Japan [19]. The averages of the JKOM were 1.1 ± 2.1 [0–8] at baseline and 2.8 ± 4.1 [0–30] at the follow-up. The detailed demographic data are shown in Table 1. The demographic data of non-responders (n = 356) were not significantly different from those of responders (data not shown).

A total of 1262 participants whose total score worsened by 3 or more points in the second survey were identified. Univariate analysis was performed to identify relevant factors and showed that greater age, female sex, higher BMI, weight increase, heavy drinking, worse mental health, and the presence of low back pain/disability were significant factors in the development of knee pain (data not shown). Multivariate analysis identified the same risk factors that were identified in univariate analysis, except for heavy drinking (Table 2, model 1). Then, multivariate analysis with stepwise selection identified six factors that were risk factors for the incidence of knee pain (Table 2, model 2). Furthermore, low back pain/disability was statistically significant even if it was treated as a continuous variable; odds ratios were 1.10 (95% CI 1.08–1.13) in univariate analysis and 1.08 (95% CI 1.06–1.11) in multivariate analysis.

Based on these analyses, we developed a predictive score (Table 3). The total possible score is 14, as outlined above, consisting of 4 points for age, 2 points for female sex, 3 points for BMI, 1 point for weight increase, 2 points for mental health, and 2 points for low back pain. The risk of developing new knee pain ranged from 11.0 to 63.2% depending on the total score (Table 4). The model calibration was good, with close agreement between the predicted and observed incidence of new knee pain (Additional file 2: Figure S1). The calculated C-index was 0.6326. The scores 2 and 4 of age, 2 of sex, 3 of BMI, 2 of mental health, 2 of low back pain, and 1 of weight increase were attributed to this population of 51.5%, 15.8%, 64.8%, 19.3%, 52.5%, 30.3%, and 54.6%, respectively. The score was normally distributed in this population (Additional file 3: Table S1).

Nevertheless, this study involves some unavoidable limitations. First, the origin of knee symptoms was not confirmed by any method. Knee symptoms may be confused with lower leg pain originating from back ailments, although the JKOM questionnaire is designed to elicit knee-specific symptoms. Second, we did not collect any imaging data, which would increase the reliability of the score. However, the purpose of the current study was to develop a reliable score that did not require any invasive measurements, and any imaging studies should be used in a different setting. Third, our analyses were performed with a predetermined set of data. We cannot exclude the existence of other factors that could contribute to the prediction of new knee symptoms, including injury history, educational level, and metabolic syndrome. Fourth, the threshold of the score determining the need for appropriate intervention is unclear, and this should be extensively studied in a future longitudinal, proof-of-concept study. Finally and importantly, prediction score constructed by regression analysis tends to overfit the data which the score was derived from, especially when using stepwise selection. Therefore, the performance of our score needs to be confirmed by external patient data.