Prediction using machine learning of cumulative drug use for neuropathic pain treatment and its determinants in people with diabetic neuropathy in a national cohort

Study type

• Study topic:
  Other
• Study topic, other:
  Prediction modelling of cumulative pharmacological use in diabetic neuropathy using real-world data.
• Study type:
  Non-interventional study
• Scope of the study:
  Drug utilisation

  Method development or testing
• Data collection methods:
  Secondary use of data

Non-interventional study

• Non-interventional study design:
  Cohort

Study drug and medical condition

• Medicinal product name, other:
  The pharmacological exposure of interest will include the following active substances and ATC codes: N02BF01 gabapentin, N02BF02 pregabalin, N06AX21 duloxetine, N06AX16 venlafaxine, N06AX23 desvenlafaxine, N06AA02 imipramine, N06AA04 clomipramine, N06AA06 trimipramine, N06AA09 amitriptyline, N06AA10 nortriptyline and N06AA12 doxepin. Some active substances were not available as selectable INN taxonomy terms in the catalogue interface and are therefore specified using their corresponding ATC codes.
• Study drug International non-proprietary name (INN) or common name:
  GABAPENTIN

  PREGABALIN

  DULOXETINE

  VENLAFAXINE

  DESVENLAFAXINE

  CLOMIPRAMINE HYDROCHLORIDE

  AMITRIPTYLINE
• Anatomical Therapeutic Chemical (ATC) code:
  (N02BF01) gabapentin

  gabapentin

  (N02BF02) pregabalin

  pregabalin

  (N06AX21) duloxetine

  duloxetine

  (N06AX16) venlafaxine

  venlafaxine

  (N06AX23) desvenlafaxine

  desvenlafaxine

  (N06AA02) imipramine

  imipramine

  (N06AA04) clomipramine

  clomipramine

  (N06AA06) trimipramine

  trimipramine

  (N06AA09) amitriptyline

  amitriptyline

  (N06AA10) nortriptyline

  nortriptyline

  (D04AX01) doxepin

  doxepin
• Medical condition to be studied:
  Diabetic neuropathy

Population studied

• Short description of the study population:
  Adults aged 18 years or older included in BIFAP with an incident recorded diagnosis of diabetic neuropathy between 1 January 2019 and 31 December 2021. The study will be conducted in Spain using routinely collected real-world data from BIFAP, mainly from longitudinal primary care electronic health records and, when available, complementary hospital information. The index date will be defined as the date of the first recorded diagnosis of diabetic neuropathy during the inclusion period, and each patient will be followed for three years from that date. Patients younger than 18 years, patients with active cancer, women pregnant during follow-up, patients with predefined active psychiatric disorders, patients with insufficient information to define essential study variables, and patients with incomplete follow-up preventing adequate longitudinal observation will be excluded. No sampling will be performed; all eligible patients available in BIFAP who meet the selection criteria will be included.
• Age groups:
  • Adult and elderly population (≥18 years)

    • Adults (18 to < 65 years)

      • Adults (18 to < 46 years)
      • Adults (46 to < 65 years)
    • Elderly (≥ 65 years)

      • Adults (65 to < 75 years)
      • Adults (75 to < 85 years)
      • Adults (85 years and over)

Study design details

• Study design:
  Retrospective longitudinal cohort study using BIFAP data. Adults with incident diabetic neuropathy recorded between 2019 and 2021 will be followed for three years from first diagnosis to model cumulative registered pharmacological use.
• Main study objective:
  To develop and validate a predictive model to estimate, from baseline patient characteristics, the three-year cumulative defined daily dose of drugs used for neuropathic pain treatment in patients with diabetic neuropathy, specifically tricyclic antidepressants, serotonin-norepinephrine reuptake inhibitors and α2δ subunit ligands.
• Setting:
  The study will be conducted in Spain using routinely collected real-world data from BIFAP. The cohort will include adults aged 18 years or older with an incident recorded diagnosis of diabetic neuropathy between 1 January 2019 and 31 December 2021. The index date will be the first recorded diagnosis of diabetic neuropathy within the inclusion period. Each patient will have an individualised three-year follow-up from the index date. Patients younger than 18 years, patients with active cancer, women pregnant during follow-up, patients with predefined active psychiatric disorders, patients with insufficient information to define essential variables, and patients with incomplete follow-up preventing adequate longitudinal observation will be excluded. No treatment arms or comparators are defined, as this is a non-interventional retrospective cohort study.
• Comparators:
  No formal comparator group is defined. This is a non-interventional retrospective cohort study. The study will compare levels and longitudinal patterns of registered pharmacological use across patients according to baseline demographic, clinical, laboratory, pharmacological and healthcare utilisation characteristics.
• Outcomes:
  The primary outcome will be three-year cumulative pharmacological use, expressed as cumulative defined daily dose, of drugs used for neuropathic pain treatment: tricyclic antidepressants, serotonin-norepinephrine reuptake inhibitors and α2δ subunit ligands. This outcome will be calculated from BIFAP prescription records, overall and by drug family, and will be interpreted as registered pharmacological use rather than actual consumption or adherence. Complementary longitudinal indicators may include treatment initiation, persistence, discontinuation, treatment escalation and switching between pharmacological families, depending on data availability and quality.
• Data analysis plan:
  Analyses will be conducted in R and Python using predefined reproducible workflows. Descriptive analyses will summarise baseline characteristics, outcome distribution, missing data and correlations among predictors. Candidate predictors will be restricted to variables available at or before the index date to avoid look-ahead bias. Data will be split into training, validation and independent test sets, with preprocessing, imputation and encoding performed within closed pipelines to avoid data leakage. The reference model will be XGBoost or gradient boosting with Tweedie loss, suitable for skewed continuous outcomes with possible excess zeros. An alternative multitask hurdle neural network may be developed to model separately the probability and magnitude of pharmacological exposure. Performance will be assessed using MAE, RMSE and R² for continuous outcomes, and AUC-ROC, precision-recall metrics, sensitivity, specificity and calibration for binary hurdle components. Interpretability will be assessed using SHAP and, where appropriate, DeepSHAP. Sensitivity analyses will evaluate the robustness of analytical decisions.