Investigating disease burden, management approaches, and clinical course prediction of low back pain using machine learning based on data from 6,426 patients in Hong Kong

Purpose: Low back pain (LBP) is the major cause of years lived with disability worldwide with largely unclear etiology. Many developed recurrent or chronic LBP, which cast a heavy burden on health systems. We hence aimed to: (1) evaluate the disease burden and management plan of LBP at general practitioner
and specialist care level; and (2) develop novel screening algorithm for detecting patients at-risk of developing chronic LBP (cLBP) for early interventions.
Methods: Medical records for patients from 2011 to 2020 were extracted from a cluster of hospitals in Hong Kong. We analysed the incidence and prevalence trends of LBP cases, patterns of analgesic prescription, operations and physiotherapy referrals, and developed machine learning models to predict the onset of
cLBP for patients completing the first year of treatment. Models were trained and validated on patient record before and after 2017, respectively. We furthered test a set of adaptive models for the same task and adopted a test-then-train validation method on incoming data arranged in chronological order.
Findings: Our search identified 6,426 patients from 2011-2020. The age-adjusted prevalence for LBP among patients from Central and Western and Southern district rose from 14.4% to 25.2% in the past decade (p = 0.0002), with 80% having non-specific diagnoses. On average, tramadol and surgeries were introduced at 1.51 years (SD = 2.20 years) and 2.11 years (SD=2.33) after the initial diagnosis. Physiotherapy referrals in the first year were associated with fewer future emergency room visits (p=0.0043) and general outpatient clinic visits (p=0.0116). Adaptive random forest model (Figure 1) out-performed nonadaptive XGBoost model (Figure 2) in identifying patients with cLBP at the first year post-diagnosis (sensitivity 95.3% v 82.0%, positive predictive value 79.2% v 62.3%).
Interpretation: Given the increasing burden of LBP in Hong Kong, opioids are commonly prescribed as the first-line treatment but demands for potent alternatives and operations will rise with high prevalence of recurrent and chronic cases. Early physiotherapy referrals are associated with favourable recovery and less emergency room visits. Our machine learning models effectively identify LBP patients at high risk of persistence for further evaluations or focused care. Adaptive models, which can adjust for dynamic data trends, were shown to be effective in delivering robust prediction despite changes in epidemiology.