PaperlayerView exact revision on arXiv

Source-linked AI summary

Random Forest-Based Prediction of Stroke Outcome

Carlos Fernandez-Lozano, Pablo Hervella, Virginia Mato-Abad, Manuel Rodriguez-Yanez, Sonia Suarez-Garaboa, Iria Lopez-Dequidt, Ana Estany-Gestal, Tomas Sobrino, Francisco Campos, Jose Castillo, Santiago Rodriguez-Yanez, Ramon Iglesias-Rey

arXiv:2402.00638v1cs.LG

TL;DR

This study addresses the challenge of predicting 3-month mortality and morbidity after stroke. It applies Random Forest models to stroke-registry data and reports high mortality-prediction performance across patient groups, while noting limits of its single-center retrospective dataset.

  • Problem

    Predicting functional outcomes after stroke is difficult, and an initial analysis using 65 features yielded AUC below 0.65 in the best case amid class imbalance and correlated features.

    Source: Discussion · p. 6, Data pre‑processing · p. 10

  • Method

    The study used Random Forest models on demographic, clinical, molecular, and neuroimaging data, selecting seven variables for prediction.

    Source: Results · p. 2, Discussion · p. 7

  • Results

    The classifier showed high performance across three groups compared with ROC analysis of seven input variables, with NIHSS(0), NIHSS(24), or NIHSS(48) among the most influential predictors.

    Source: Discussion · p. 8

  • Takeaways & Limitations

    Random Forest can be used to predict long-term mortality and morbidity in stroke patients, with NIHSS at 24 and 48 hours and axillary temperature among the important variables.

    Source: Conclusions · p. 9

    This was a retrospective, single-center study with a relatively small clinical dataset; multicenter and prospective validation is needed.

    Source: Discussion · pp. 8–9

Abstract

from arXiv · show

We research into the clinical, biochemical and neuroimaging factors associated with the outcome of stroke patients to generate a predictive model using machine learning techniques for prediction of mortality and morbidity 3 months after admission. The dataset consisted of patients with ischemic stroke (IS) and non-traumatic intracerebral hemorrhage (ICH) admitted to Stroke Unit of a European Tertiary Hospital prospectively registered. We identified the main variables for machine learning Random Forest (RF), generating a predictive model that can estimate patient mortality/morbidity. In conclusion, machine learning algorithms RF can be effectively used in stroke patients for long-term outcome prediction of mortality and morbidity.

Results

Among the 4,922 ischemic stroke (IS) and 1,100 intracerebral hemorrhage (ICH) patients, the combined-group Random Forest model was the most stable for mortality prediction. Mortality and morbidity prediction varied by stroke group, with NIHSS measures among the most influential variables.

Results

  • Among 4,922 IS and 1,100 ICH patients, poor 3-month outcomes affected 47.5% and 58.6%, respectively.IS morbidity and mortality were 33.4% and 13.2%; ICH morbidity and mortality were 27.6% and 30.2%.Source: Results · p. 2
  • Seven selected inputs were NIHSS(0), NIHSS(24), NIHSS(48), T(0), ED, LEU(0), and GLU(0); NIHSS(48) and NIHSS(24) led mortality importance across groups.ED, T(0), and NIHSS(0) were also important in IS and ICH models; LEU(0) and GLU(0) helped balance results in the combined model.Source: Results · p. 2, Prediction of mortality · p. 4

Prediction of mortality.

  • 0.904 ± 0.025 AUC and 0.825 ± 0.030 accuracy (ACC) marked the most stable IS + ICH mortality model; IS reached 0.909 ± 0.032 AUC and 0.833 ± 0.040 ACC.ICH predictions were less stable: 0.875 ± 0.048 AUC and 0.8 ± 0.052 ACC across 100 repetitions, with AUC ranging from 0.7104 to 0.9837.Source: Prediction of mortality · p. 4, Figure 2 · p. 5

Prediction of morbidity.

  • For morbidity, NIHSS(48) and NIHSS(24) ranked highest overall; ED mattered for IS and combined cases, GLU(0) for ICH, and NIHSS(0) had differing effects by group.NIHSS(0) worsened IS classification but was a key ICH variable.Source: Prediction of morbidity · p. 5
  • Morbidity prediction AUC/ACC were 0.738/0.683 for IS and 0.755/0.700 for IS + ICH, versus 0.667/0.618 for ICH.The authors reported no major differences between the IS and IS + ICH groups.Source: Prediction of morbidity · p. 5

AUC with 7 input variables.

  • For individual-input AUCs, NIHSS(24) and NIHSS(48) led both outcomes in IS and combined groups; ICH morbidity favored NIHSS(0) and NIHSS(24), while ICH mortality favored NIHSS(24) and NIHSS(48).For IS + ICH and IS, the reported individual-variable AUCs were higher for NIHSS(48) than NIHSS(24) in both morbidity and mortality prediction.Source: AUC with 7 input variables · p. 5

Discussion

The study applies machine learning to 3-month stroke outcomes, addressing limited prior work across stroke types and varied prediction methods. Its models performed well for mortality and more modestly for morbidity, but the single-center retrospective design and other scope constraints warrant further validation.

Discussion

  • Accurate outcome prediction could inform long-term decisions and individualized rehabilitation planning, while machine learning may capture nonlinearities and interactions among predictors.The authors describe potential relevance to patient treatment, health-center organization, and domestic conditions.Source: Discussion · p. 6
  • Mortality AUC exceeded 0.90 across IS, ICH, and combined groups, while morbidity AUC exceeded 0.75 in all three groups.The IS group had the best mortality results; NIHSS (24) and NIHSS (48) were the most relevant variables across groups.Source: Discussion · p. 7
  • NIHSS (24) and NIHSS (48) were the leading predictors; ED was relevant for IS morbidity and GLU (0) for ICH morbidity.The study also found NIHSS (0) important for ICH, and identified axillary temperature and ED as variables not to disregard.Source: Discussion · p. 7
  • The study used demographic, clinical, molecular, and neuroimaging variables in a large registry, examining IS and ICH together and separately.The authors note that earlier studies varied in sample size, patient groups, variables, training protocols, and models.Source: Discussion · pp. 6–7
  • Compared with ROC analysis of the seven input variables, the machine-learning classifier performed well across three groups, with NIHSS measures among the most influential predictors.NIHSS (0), NIHSS (24), or NIHSS (48) was most influential, depending on the group.Source: Discussion · p. 8
  • The findings are limited by a retrospective, single-center design, relatively small clinical dataset, unbalanced stroke groups, use of RF alone, and lack of patient subgroup stratification.The authors call for multicenter prospective training and validation to verify the model and selected variables.Source: Discussion · pp. 8–9

Conclusions

Random Forest can be used to predict long-term mortality and morbidity in stroke patients. The authors identify NIHSS at 24 and 48 hours and axillary temperature as important variables for 3-month outcome prediction.

Conclusions

  • Random Forest can be effectively used to predict long-term mortality and morbidity in stroke patients.The authors suggest future studies could incorporate imaging and genetic information.Source: Conclusions · p. 9
  • NIHSS at 24 and 48 hours and axillary temperature are important variables to consider when predicting patients’ 3-month evolution.The authors also suggest the model could be used in other scopes with similar data, such as traumatic brain injury or dementia.Source: Conclusions · p. 9

Materials and methods

The study used a prospectively registered stroke cohort, measured clinical and neuroimaging variables, and modeled 3-month mortality and morbidity across ischemic stroke and intracerebral hemorrhage groups. Random Forest models were evaluated with repeated undersampling and cross-validation after feature selection.

Patient selection.

  • The analysis included patients with ischemic stroke and intracerebral hemorrhage, excluding those who died in the first 24 hours or lacked 3-month follow-up.The data were analyzed retrospectively for September 2007 to September 2017.Source: Patient selection · p. 9

Demographic, clinical, molecular and neuroimaging variables.

  • The registry captured demographics, medical history, vital signs, laboratory results, NIHSS assessments, functional outcomes, and neuroimaging measures.CT was performed in all patients, while MRI was performed in selected patients; ICH and perihematomal edema volumes used the ABC/2 method.Source: Demographic, clinical, molecular and… · p. 9

Outcome endpoints.

Machine learning.

  • Random Forest aggregated predictions and provided feature-importance measures, including Gini importance, for the stroke outcome models.The method’s feature importance was calculated using the Gini impurity index [37–39].Source: Machine learning · pp. 9–10

Data pre‑processing.

  • Class imbalance was addressed with random undersampling, evaluated over 100 repetitions of tenfold cross-validation, and filter selection reduced the feature set from 65 variables.Filter scores used a training-data T-test, with low-scoring features removed at a manual cutoff.Source: Data pre‑processing · p. 10

Statistical analysis.

  • Model performance was measured by mean ± SD AUC across tenfold validation, and ROC analysis compared logistic regression with machine-learning models using seven inputs.The seven selected inputs were compared as potential 3-month morbidity and mortality clinical markers.Source: Statistical analysis · p. 10

Data availability

The manuscript states that its data are available in the text, with further anonymized data potentially available to qualified investigators upon reasonable request.

Data availability.

  • Study data are reported as available within the manuscript, and further anonymized data may be requested by qualified investigators.The additional data are conditional on a reasonable request.Source: Data availability · p. 10

Funding

The study reports support from Spanish national and regional bodies, research networks, and European funding programs.

Funding.

  • The authors acknowledge grants and research-contract support from Spanish ministries, Xunta de Galicia, ISCIII, RETICS-INVIC-TUS PLUS, and European Union FEDER programs.The listed support also includes the Miguel Servet Program and several regional and institutional initiatives.Source: Funding · p. 12

Additional information

The paper provides supplementary material information, correspondence contacts, publisher notices, and its open-access license terms.

Additional information.

  • The online article has supplementary material, and correspondence requests are directed to S.R.-Y. or R.I.-R.Reprints and permissions information is available from the publisher.Source: Additional information · p. 12
  • The article is licensed under Creative Commons Attribution 4.0, subject to attribution and the stated conditions for third-party material.The publisher states neutrality regarding jurisdictional claims in maps and affiliations.Source: Additional information · p. 12
Loading 2402.00638v1…