INTRODUCTION
Intervertebral disc (IVD) degeneration (IDD), a leading cause of low back pain (LBP), has biological and mechanical factors compromising disc integrity [1]. Despite its prevalence, early detection, treatment strategies, and the complex aetiology of LBP remain limited, highlighting the need for data-driven clinical decision support systems.
Visualization-based decision support tools have proven effective in simplifying complex biomedical data, reducing medical errors, and enhancing healthcare efficiency. The Disease State Index (DSI) [2] computes a scalar value measuring the variable alignment with a patient’s LBP or IDD state. This study employs the Disease State Fingerprint (DSF) technique [3] to visualize biopsychosocial [4] relationships with LBP and IDD.
METHODS
Biopsychosocial variables were analyzed from a subset of the Northern Finland Birth Cohort (NFBC1966), comprising 5,525 individuals born in 1966 [5] and assessed at 46 years old at follow-up. The dataset featured heterogeneous variables—scalars, categorical data, and clinical reports—along with missing values, all effectively handled by the DSF.
The study focused on LBP and IDD, with two classes defined for each: controls and positives. LBP intensity (n=4531) was assessed via Numerical Rating Scale (NRS 0–10) with NRS ≥ 5 indicating positive cases. IDD (n=1,410) was classified using Pfirrmann grades from MRI across the five lumbar segments, with grades ≥3 denoting positivity.
A total of 673 biopsychosocial variables were analyzed, covering demographics (e.g., gender, daily habits, economic status, psychological traits), laboratory results (e.g., cortisol, vitamin levels, blood tests), musculoskeletal assessments, and anthropometrics. MRI-based features such as Pfirrmann grades, Modic changes, and radiomics from T2-weighted lumbar scans were added.
The six models were trained using an 80/20 training-testing split with 10-fold cross-validation.
RESULTS
Figure 1A showed strong discriminative performance across all models (AUC > 0.8), particularly for lower lumbar levels. Figure 1B demonstrates the DSF and DSI’s capability to distinguish controls from positives for pain Intensity and IDD. Testing accuracy exceeded 70%, with LBP, L4-L5, and L5-S1 models achieving >80%, aligning closely with validation accuracy (Table 1).
For LBP (Fig. 1C), key predictors included disability, difficulty in sleep duration, depression, fear of chronic pain (DSI > 0.7), and elevated cortisol levels (DSI ~ 0.6). L4-L5 and L5-S1 IDD levels were significant for pain occurrence but less for intensity.
The five IDD levels shared similar DSF tree structures, emphasizing level-specific features. In L5-S1 (Fig. 1D), radiomics features (DSI > 0.7), MRI-assessed Modic changes, and Schmorl’s nodes (DSI > 0.6) were critical, alongside the questionnaires (DSI > 0.6) with disability, low physical activity, smoking, and osteoarthritis.
DISCUSSION
The models accurately differentiated controls from positives, demonstrating IDD's role in LBP, while intensity strongly correlated with psychosocial factors. Radiomics and Modic changes emerged as robust predictors for IDD classification. Interestingly, sex was insignificant in LBP and IDD evaluations.
DSF holds promise for stratifying pain by IDD phenotypes and psychosocial factors. Future work will integrate patient-specific finite element simulations via the Pixel2Mechanics [6] tool, leveraging longitudinal NFBC1966 data to uncover mechanistic insights [7] into IDD and pain for early detection.
FOUNDING
MSCA- 2020-ITN-ETN GA: 955735, ERC-2021-CoG-O-Health-101044828), and ICREA Acadèmia programme.