Decision Support in Spine Treatment Guided by MachineLearning and Registry Data

Casper Friis Pedersen

Research output: ThesisPh.D. thesis

11 Downloads (Pure)


Lumbar spinal stenosis is an abnormal narrowing of the spinal canal that causes impingement of thenerves traveling through the lower back into the legs. Symptoms may include pain, weakness ornumbness in legs, calves or buttocks. Walking, standing and extending the lower back can aggravatesymptoms. Lumbar spinal stenosis is a frequent age-related spinal disorder with an estimatedprevalence of 20-50% among the geriatric segment of the population and is the leading cause ofspinal surgery among the elderly. According to the Danish national spine registry DaneSpine, 3/4 ofpatients can expect considerable pain relief 1 year after surgery and 2/3 will experienceimprovements in quality of life. About 1/4 - 1/3 of patients do not feel a clear improvement.

The variation in surgical outcome often makes it difficult to communicate a reliable prognosis to thepatient. The main purpose of this thesis was to establish if reliable individualized estimates ofoutcome could be computed preoperatively through predictive algorithms modelled on existing registry data.

In study 1, we found that the predictive performance of the Swedish decision support DialogueSupport did not generalize well to Danish patient samples. While AUC values were comparablewith the results reported by the authors of the Dialogue Support, both calibration plots andperformance metrics revealed a low ability to correctly identify unfavourable outcomes (true negatives).

In study 2, seven different machine learning algorithms were applied to Danish spine data. Onaverage, they performed nearly equally well but variation was found across outcome measures. TheEQ-5D and VAS back models performed almost equally well, while the ODI and VAS leg modelswere less convincing. VAS leg and Return to work models exhibited the largest differences inperformance between algorithms. MARS and deep learning performed consistently well.

In study 3, non-operated LSS patients were matched with operated patients to ensure equivalentbaseline characteristics. Both groups were diagnosed with MRI-confirmed LSS by spine surgeons.The outcome was compared at 1 year following their initial consultation with a spine surgeon.Although both groups improved on average, differences were in favour of the operated patientswhether measured as mean improvement or proportions reaching a minimal clinical importantdifference. Less than half of the non-operated achieved MCID on EQ-5D, VAS back/legs compared to2/3 of the operated.

In study 4, we developed a decision support tool PROPOSE capable of predicting personalizedoutcome at 1 year following surgery. It was validated by surgeons in a clinical consultation setting.The predictive performance of PROPOSE with respect to EQ-5D, VAS back and ODI was fair toexcellent, while VAS leg was less predictive in comparison. The tool could prove useful as a mean tolet patients and surgeons engage in discussions on likely outcome and support clinical decisionmaking. Before considering its use in clinical practice it should however be thoroughly validated onexternal data sources independent of the data used in development.
Original languageEnglish
Awarding Institution
  • University of Southern Denmark
  • Andersen, Mikkel Østerheden, Principal supervisor
  • Carreon, Leah, Co-supervisor
  • Eiskjaer, Søren, Co-supervisor, External person
Date of defence26. May 2023
Publication statusPublished - 11. May 2023


  • machine learning
  • prediction
  • decision-making
  • spinal stenosis
  • lumbar herniated disc


Dive into the research topics of 'Decision Support in Spine Treatment Guided by MachineLearning and Registry Data'. Together they form a unique fingerprint.

Cite this