Preoperative Prognosis Assessment of Lumbar Spinal Surgery for Low Back Pain and Sciatica Patients based on Multimodalities and Multimodal Learning

Li-Chin Chen, Jung-Nien Lai, Hung-En Lin, Hsien-Te Chen, Kuo-Hsuan Hung, Yu TsaoarXiv 2023 · DOI · arXiv

Introduction

Lumbar spinal surgery is a high-stakes intervention for chronic low back pain and sciatica. While many patients find relief, a significant number—up to 39% in some studies—do not experience meaningful improvement, a phenomenon often referred to as "failed back surgery syndrome." This paper introduces a multimodal machine learning framework to predict the success of lumbar spinal surgery before a patient ever goes under the knife. By combining standard clinical metrics with Traditional Chinese Medicine (TCM) assessments, surgical plan descriptions, and voice recordings, the researchers achieved a prediction accuracy of 81%, offering a potential tool to reduce the incidence of surgical failure.

Problem Statement

Currently, surgeons lack a reliable, objective way to evaluate surgical outcomes in advance. Existing tools like the North American Spine Society instrument haven't gained widespread traction because they often fail to capture the complex, heterogeneous nature of patient health. This paper addresses the need for a "lightweight" preoperative assessment tool that integrates diverse data types to provide a more holistic prognosis.

Methodology and Data Integration

The researchers developed a multimodal learning pipeline that processes four distinct types of data collected before surgery:

  1. Tabular Data: Standard Western clinical metrics including age, BMI, pain scales (VAS), disability indices (ODI), and ASA physical status.
  2. TCM Body Constitution: Scores for "Yang-Xu" (deficiency), "Yin-Xu," and "Stasis," representing physiological state from an Eastern medicine perspective.
  3. Free Text: Descriptions of the planned surgical approach (e.g., minimally invasive vs. open fusion).
  4. Audio: Recordings of patients pronouncing five vowels (/a/, /e/, /i/, /o/, /u/), based on the TCM theory that voice reflects internal energy balance.
Overview of the data flow, including preoperative inputs and the calculation of ground truth labels from surgical and post-operative outcomes.

Technical Architecture

The system utilizes specialized unimodal encoders to extract features from each data type:

  • Tabular Encoders: The team compared tree-based models (CatBoost) with 1D Convolutional Neural Networks (CNNs).
  • Text Encoder: Employs PubMedBERT fine-tuned with a Long Short-Term Memory (LSTM) network to process surgical descriptions.
  • Acoustic Encoder: Uses Short-Time Fourier Transform (STFT) followed by an LSTM to analyze vocal characteristics.
Detailed unimodal architectures for tabular, text, and acoustic data processing, highlighting normalization, embedding, and feature extraction layers.

To combine these modalities, three fusion strategies were evaluated: Early Fusion (EF) (joining features into a single vector), Joint Fusion (JF) (joining learned representations within the neural network), and Late Fusion (LF) (averaging prediction probabilities).

Dataset and Ground Truth

The study utilized a custom, prospective cohort dataset of 105 patients diagnosed with low back pain or sciatica at China Medical University Hospital, Taiwan. A binary ground truth label ("desirable" vs. "undesirable" prognosis) was established by calculating a composite score from eight post-surgical outcomes, including pain reduction, blood loss, and complications.

Key Results and Insights

  • Top Performance: The combination of Tabular + Text data using CatBoost in an Early Fusion configuration achieved the highest performance, with an accuracy of 0.81 and an AUROC of 0.829.
  • The Power of the Surgical Plan: Surprisingly, the free-text description of the surgical approach was a dominant predictor. Using NLP to extract features from how a surgeon intends to operate provided more signal than many traditional numerical scales.
  • Eastern & Western Synergy: Integrating TCM body constitution assessments helped capture holistic patient health. For example, "Stasis" scores correlated with audio features, suggesting a link between internal physiological state and vocal characteristics.
  • Interpretability: By using Integrated Gradients (IG), the authors showed that while text was the strongest predictor, age and ASA physical status remained critical anchors for the model's decision-making.
  • Audio Contribution: While audio was a novel addition, it did not significantly boost accuracy on its own (0.607 accuracy). However, it served as an auxiliary modality that improved the generalizability of the joint models.

Conclusion and Future Work

This study demonstrates that a multimodal approach, incorporating both traditional clinical data and novel inputs like TCM assessments and acoustics, can significantly improve preoperative prognosis for spinal surgery. The primary limitation remains the small sample size (105 patients), which favored tree-based models over complex deep learning for tabular data. Future work will focus on expanding the dataset and exploring more sophisticated pre-trained models for audio to further refine this clinical decision-support tool.

Summary updated

Abstract

Low back pain (LBP) and sciatica may require surgical therapy when they are symptomatic of severe pain. However, there is no effective measures to evaluate the surgical outcomes in advance. This work combined elements of Eastern medicine and machine learning, and developed a preoperative assessment tool to predict the prognosis of lumbar spinal surgery in LBP and sciatica patients. Standard operative assessments, traditional Chinese medicine body constitution assessments, planned surgical approach, and vowel pronunciation recordings were collected and stored in different modalities. Our work provides insights into leveraging modality combinations, multimodals, and fusion strategies. The interpretability of models and correlations between modalities were also inspected. Based on the recruited 105 patients, we found that combining standard operative assessments, body constitution assessments, and planned surgical approach achieved the best performance in 0.81 accuracy. Our approach is effective and can be widely applied in general practice due to simplicity and effective.
BibTeX
@misc{chen2023preoperative,
  author = {Li-Chin Chen and Jung-Nien Lai and Hung-En Lin and Hsien-Te Chen and Kuo-Hsuan Hung and Yu Tsao},
  title = {{Preoperative Prognosis Assessment of Lumbar Spinal Surgery for Low Back Pain and Sciatica Patients based on Multimodalities and Multimodal Learning}},
  year = {2023},
  eprint = {2303.09085},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2303.09085},
}
Figure