Machine learning for cerebral blood vessels' malformations

Cerebral aneurysms and arteriovenous malformations are life-threatening hemodynamic pathologies of the brain. While surgical intervention is often essential to prevent fatal outcomes, it carries significant risks both during the procedure and in the postoperative period, making the management of these conditions highly challenging. Parameters of cerebral blood flow, routinely monitored during medical interventions or with modern noninvasive high-resolution imaging methods, could potentially be utilized in machine learning-assisted protocols for risk assessment and therapeutic prognosis. To this end, we developed a linear oscillatory model of blood velocity and pressure for clinical data acquired from neurosurgical operations. Using the method of Sparse Identification of Nonlinear Dynamics (SINDy), the parameters of our model can be reconstructed online within milliseconds from a short time series of the hemodynamic variables. The identified parameter values enable automated classification of the blood-flow pathologies by means of logistic regression, achieving an accuracy of 73 %}. Our results demonstrate the potential of this model for both diagnostic and prognostic applications, providing a robust and interpretable framework for assessing cerebral blood vessel conditions.

💡 Research Summary

**
The paper addresses the pressing clinical problem of diagnosing and prognosing cerebral arterial aneurysms (AA) and arteriovenous malformations (AVM), both of which carry high morbidity and mortality. While surgical intervention is often required, the decision-making process is hampered by uncertainty regarding the risk of rupture and the outcome of treatment. The authors propose a data‑driven approach that leverages intra‑operative measurements of blood velocity v(t) and pressure p(t) to build a compact, interpretable dynamical model that can be identified in real time and subsequently used for classification of pathology.

Model Development and Sparse Identification
Previous work modeled the pressure dynamics with a nonlinear Liénard‑type second‑order differential equation containing polynomial functions A(p) and B(p). The authors argue that this formulation is overly complex for real‑time clinical use, primarily because it requires high‑dimensional optimization and is sensitive to initial guesses. To overcome these limitations, they construct an extensive library Θ of candidate functions, including powers of p, its first derivative ˙p, cross‑terms, and the forcing velocity v(t). Using the Sparse Identification of Nonlinear Dynamics (SINDy) framework together with Sequentially Thresholded Least Squares (STLS), they perform a sparsity‑threshold sweep (η = 0.1, 1.0, 5.0). At η = 5.0 all nonlinear terms are eliminated, yielding a simple forced damped harmonic oscillator:

  ¨p + a ˙p + b p = ε v.

Only three parameters (a, b, ε) remain, dramatically reducing computational load and enabling parameter estimation within milliseconds from a short (≈5 cardiac cycles) time series.

Data Acquisition and Pre‑processing
The dataset consists of ten patients (five AA, five AVM) operated at the Meshalkin Research Institute in Novosibirsk. Velocity and pressure were recorded via a 0.34 mm intravascular Doppler guidewire at 200 Hz, yielding 3.3–5.4 s recordings per patient (Δt = 5 ms). After detrending by subtracting the mean, the authors compute first and second derivatives using central finite differences (second‑order accurate) and apply a low‑pass filter to remove high‑frequency noise.

Parameter Robustness
Across all patients, the identified parameters exhibit a relative deviation of at most 24 %, indicating that the linear model captures the dominant dynamics despite inter‑patient variability. The root‑mean‑square error (RMSE) between measured and simulated pressure trajectories is low, confirming good fit quality.

Classification Using Logistic Regression
The three identified parameters serve as a compact feature vector for each patient. A multinomial logistic regression model is trained to discriminate among three classes: (0) AA flow, (1) AVM flow, and (2) post‑surgical normal flow. Despite the limited sample size, the classifier achieves an overall accuracy of 73 %, which the authors deem promising given the modest data volume and the simplicity of the feature set.

Strengths

1. Real‑time feasibility – SINDy with a high sparsity threshold reduces the identification problem to solving an over‑determined linear system, eliminating the need for iterative nonlinear optimization.
2. Interpretability – The final model is a classic damped oscillator, allowing clinicians to relate parameters to physical concepts such as vascular compliance (a) and stiffness (b).
3. Dimensionality reduction – From hundreds of raw time‑series points per patient, the method extracts only three meaningful descriptors, facilitating downstream machine‑learning tasks.

Limitations

• Sample size – Ten patients limit statistical power and raise concerns about overfitting. External validation on larger, multi‑center cohorts is needed.
• Invasive measurement – The study relies on intravascular Doppler data; translation to non‑invasive modalities (e.g., phase‑contrast MRI) remains to be demonstrated.
• Model comparison – Only logistic regression is evaluated; more sophisticated classifiers (SVM, neural networks) could potentially improve accuracy.
• Physiological interpretation of ε – While ε scales the forcing term, its direct physiological meaning (e.g., vascular resistance) is not fully explored.

Future Directions
The authors suggest expanding the dataset, incorporating non‑invasive imaging, and exploring higher‑tier machine‑learning pipelines that use the identified parameters as inputs. Additionally, integrating patient‑specific anatomical information (e.g., vessel geometry from imaging) could refine the physical interpretation of the parameters and improve prognostic power.

Conclusion
This work demonstrates that a sparse, linear dynamical model of cerebral blood‑pressure dynamics can be identified in real time from intra‑operative measurements and that the resulting parameters are sufficiently informative to classify major cerebrovascular pathologies with respectable accuracy. The approach balances computational efficiency, interpretability, and clinical relevance, offering a promising foundation for decision‑support tools in neurosurgery.