Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Predicting Postoperative Stroke in Elderly SICU Patients: An Interpretable Machine Learning Model Using MIMIC Data

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A CatBoost model trained on the first 24 hours of routine ICU data predicts in-hospital stroke in elderly SICU patients with an AUROC of 0.8868, and names cerebrovascular disease, creatinine, and systolic blood pressure as the dominant…

desk verdict The ML pipeline is competent, but the outcome label is not temporally anchored, so the reported AUROC likely reflects billing-code history rather than postoperative stroke. read the letter →

arxiv 2506.03209 v1 pith:WLJWRURE submitted 2025-06-02 q-bio.QM cs.AIcs.LG

Tinghuan Li , Shuheng Chen , Junyi Fan , Elham Pishgar , Kamiar Alaei , Greg Placencia , Maryam Pishgar This is my paper
classification q-bio.QMcs.AIcs.LG
keywords postoperativestrokeelderlySICUpatientsmachinelearningCatBoostMIMIC-IIIMIMIC-IVSHAPclinicalriskprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Postoperative stroke is a rare but devastating complication in elderly surgical ICU patients, and this paper tries to show that a machine learning model built only from the first 24 hours of routine ICU data can flag who is at risk. Using 19,085 MIMIC-III and MIMIC-IV admissions, the authors reduced 80 candidate variables to 20 and trained eight classifiers. They report that CatBoost performs best, with AUROC 0.8868, accuracy 91.12%, and specificity 98.17%, and that cerebrovascular disease, serum creatinine, and systolic blood pressure are the strongest predictors. If the result holds, hospitals could get an early, interpretable warning system from data they already collect, without waiting for the stroke to happen.

What carries the argument

The load-bearing mechanism is CatBoost, a gradient-boosted tree algorithm whose ordered boosting is designed to avoid target leakage and which natively handles categorical and missing data, combined with a two-stage feature selection (RFECV intersected with SHAP) that narrows 80 variables to 20. The pipeline also uses iterative SVD imputation, z-score normalization, one-hot encoding, and ADASYN oversampling confined to training folds, so the reported performance is the result of the whole preprocessing-plus-classifier stack rather than any single component.

What would settle it

Retrain the model with the outcome redefined as stroke codes first appearing after ICU admission (and after surgery when procedure timing is available), or with patients who already carry a cerebrovascular-disease diagnosis excluded. If AUROC drops substantially or cerebrovascular disease falls out of the top predictors, the original claim is largely an artifact of outcome-label leakage; if performance holds, the finding survives.

Watch

Extended reading notes

Core claim

The paper's central claim is that in-hospital stroke in elderly SICU patients can be predicted, within the first 24 hours of ICU admission, by a CatBoost model using 20 clinically interpretable features, with an AUROC of 0.8868 (95% CI 0.8802–0.8937) and a specificity of 0.9817 that keeps false alarms low. The model identifies prior cerebrovascular disease, serum creatinine (minimum and maximum), and mean systolic blood pressure as the dominant risk factors, and ablation experiments show that removing these features causes the largest drops in AUROC. The authors frame this as an advance over earlier MIMIC-based stroke models and over traditional logistic-regression scores, because the gradient-boosting model captures nonlinear interactions while SHAP values keep the predictions explainable.

Load-bearing premise

The central load-bearing assumption is that a patient is counted as having postoperative stroke whenever ICD-9 codes 430, 431, 432, 433, 434, or 436 appear anywhere in the record, with no requirement that the stroke occurred after surgery or after ICU admission; if many of those codes reflect past strokes, the model may be predicting a documented history rather than a new event.

Editorial extensions

If this is right

  • A deployed version could run at ICU admission, using only the first 24 hours of data, and generate a stroke-risk alert before clinical deterioration is obvious.
  • The high specificity (0.9817) and negative predictive value (0.9209) mean most low-risk patients would be correctly cleared, so nursing and specialist resources could concentrate on the small flagged group.
  • The short, clinically named feature list (age, vitals, creatinine, BUN, PT, and seven comorbidities) makes the model auditable by clinicians and potentially portable across EHR systems.
  • The model outperforms the previously reported XGBoost baseline (AUROC 0.78) on the same MIMIC source, suggesting the preprocessing and feature-selection choices add real predictive value.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the outcome labels are not anchored to events after ICU admission, the top predictor 'cerebrovascular disease' may partly mirror the label itself; a fair test would exclude prevalent stroke histories or require new stroke codes after admission.
  • The authors did not compare against standard clinical scores like CHA2DS2-VASc in the same cohort, so it remains an open question whether CatBoost adds discrimination beyond a simple score; that comparison is a natural next experiment.
  • Because the model only uses first-24-hour statistics, it cannot capture later deterioration; integrating time-series vitals or medication events could either improve AUROC or reveal that the real signal is already present on day one.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper develops a CatBoost-based machine learning pipeline to predict 'postoperative stroke' in elderly SICU patients from MIMIC-III and MIMIC-IV. The authors construct a combined cohort of 19,085 admissions, apply a preprocessing pipeline with SVD imputation and ADASYN, select 20 features via a two-stage RFECV/SHAP procedure, and report a test AUROC of 0.8868 (95% CI 0.8802–0.8937), with cerebrovascular disease, serum creatinine, and systolic blood pressure as the leading predictors. The central claim is that this model provides an early-warning tool from routine first-24-hour ICU data.

Significance. If the central claim were valid, the paper would offer a useful, interpretable early-risk-stratification tool built on publicly available critical care data, with a clear multi-model comparison and interpretability analyses. The authors deserve credit for using the MIMIC databases, describing a detailed pipeline, reporting confidence intervals, and attempting both SHAP and ablation analyses. However, the validity of the claim hinges on three load-bearing methodological points: the outcome definition must capture incident postoperative stroke, the top feature must not encode the outcome label, and feature selection must not leak test information. As detailed below, all three points are problematic, so the reported AUROC and the clinical interpretation do not currently establish predictive utility for postoperative stroke.

major comments (4)
  1. [Patient Selection, Algorithm 1 Step 5] The outcome is defined as the presence of any ICD-9 code in {430, 431, 432, 433, 434, 436} in the MIMIC diagnoses tables, with no timestamp requirement relative to surgery or ICU admission. MIMIC diagnostic records are discharge-level billing diagnoses and include pre-existing or historical conditions. A patient admitted with a prior stroke and no new perioperative event is therefore counted as a 'postoperative stroke.' This directly contradicts the stated objective of predicting in-hospital postoperative stroke and likely explains the implausibly high stroke incidence of 13.5% (2,568/19,085). The outcome definition must be anchored to an event occurring after surgery and during the index hospitalization, for example by linking to procedure and admission timestamps.
  2. [Feature Selection and Table 1] The top predictive feature, 'cerebrovascular disease,' is a comorbidity flag derived from ICD-9 code ranges that overlap substantially with the outcome codes 430–434 and 436. Without an explicit statement of the exact ICD-9 ranges used for the comorbidity feature, the leading predictor may simply be an alternative expression of the outcome label. This would inflate the AUROC and make the SHAP and ablation conclusions about cerebrovascular disease circular. The authors must report the exact code definitions for this comorbidity and demonstrate that all outcome-defining codes are excluded from it.
  3. [Algorithm 1 Steps 4–5 and 'Feature Selection' section] The two-stage feature selection procedure (RFECV plus SHAP) is described in Algorithm 1 as Step 4, before Step 5 ('Model Development') and before the train/test split used for validation. If the 20 retained features are chosen using the full dataset, the test labels inform the selection, and the reported test AUROC is optimistically biased. The description in the Feature Selection section does not state that selection is nested inside the cross-validation loop. The authors must move feature selection into the training folds only, or at minimum demonstrate that the selected features are stable when selection is performed on training data alone.
  4. [Table 3 vs. Figure 6 and accompanying text] The reported associations in Table 3 contradict the SHAP narrative. Table 3 shows that stroke patients had significantly lower mean and minimum creatinine, lower mean heart rate, and higher mean systolic blood pressure (all p<0.0001), whereas the Figure 6 discussion states that elevated creatinine, elevated heart rate, and mean SBP contribute to higher predicted risk. This internal inconsistency undermines the interpretability claims and suggests that the model's feature effects are not stable or that the outcome definition is capturing a different population than the one described. The authors need to reconcile these results or provide SHAP dependence plots that clarify the actual directions.
minor comments (4)
  1. [Patient Selection] The definition of SICU admissions is internally inconsistent: the text lists 'MICU' as a surgical unit for MIMIC-III, which is a medical ICU, and this should be clarified or corrected.
  2. [Algorithm 1, line 9] The preprocessing step reads 'Remove features with > 20%' and is an incomplete sentence; it should specify 'remove features with missingness greater than 20%.'
  3. [Table 2] Several reported standard deviations are implausible (e.g., dbp_mean 45.68 and 101.23, sbp_max 13.54 and 25.89, temp_max 17.62 and 15.23 in the two columns), suggesting possible unit errors or merging artifacts; these values should be verified.
  4. [Model Development and Evaluation] The paper does not report the final CatBoost hyperparameters, the ADASYN sampling parameters, or the exact bootstrap procedure for confidence intervals, which limits reproducibility; a code or data-availability statement would also be helpful.

Circularity Check

2 steps flagged · score 6.0 of 10

Outcome label and top feature are the same ICD-9 cerebrovascular codes, and feature selection occurs before the train/test split, so the reported AUROC partly reflects coding duplication and label-informed feature choice rather than a clean prediction of postoperative stroke.

  1. self definitional [Patient Selection (Step 5); Feature Selection, Comorbidities; Tables 1 and 4]
    "Stroke status was determined using International Classification of Diseases, Ninth Revision (ICD-9) diagnosis codes. Patients were classified as stroke-positive if any of the following codes were present: 430, 431, 432, 433, 434, or 436."

    The outcome is defined by cerebrovascular ICD-9 codes 430-436, and the same administrative diagnosis source is used to build the comorbidity feature 'cerebrovascular disease' that is selected into Table 1 and later identified by SHAP and ablation as the most influential predictor. ICD-9 430-436 are themselves cerebrovascular disease codes, so a patient who triggers the label through one of these codes also sets the top feature to 'present.' The model can therefore separate classes by reading the same discharge diagnosis twice: once as the outcome and once as the leading feature. The reported AUROC of 0.8868 does not establish that the model is predicting an incident postoperative stroke; it may simply be detecting that the label code and the feature code co-occur by construction.

  2. fitted input called prediction [Algorithm 1, Steps 4-5; Feature Selection section]
    "Step 4: Feature Selection Apply RFECV using Random Forest ... Identify intersection with SHAP-based top features Retain 20 shared features for model training Step 5: Model Development ... 5-fold cross-validation on training data with 20 selected features"

    RFECV and SHAP feature selection are listed as Step 4, before the train/test split and before Step 5's cross-validation. The paper states that only ADASYN and preprocessing statistics are confined to training folds; it does not state that feature selection was performed inside each training fold. Thus the 20 features are selected with access to test-set labels, and the later 'independent test cohort' performance is partly a re-description of a feature set that was chosen using those labels. This is selection leakage: the model is fitted to the selection outcome as well as to the training data, so the test AUROC of 0.8868 is not a fully out-of-sample prediction.

full rationale

The paper's central claim is that CatBoost predicts in-hospital postoperative stroke in elderly SICU patients using first-24-hour ICU data, with an independent test AUROC of 0.8868. However, two constructional issues make the claim partially circular. First, the outcome is defined solely by the presence of ICD-9 codes 430, 431, 432, 433, 434, or 436 in the MIMIC diagnosis records, with no temporal anchoring to surgery or ICU admission. The feature set then includes 'cerebrovascular disease' as a comorbidity, which is derived from the same ICD-9 cerebrovascular disease code family. Because the label and the top feature are both set by the same billing diagnosis, the model can achieve high discrimination by exploiting this code-level overlap rather than by learning a genuine physiologic early-warning signal. The paper does not define the comorbidity feature with a distinct code list, and no step in the pipeline separates prior cerebrovascular disease from the outcome codes, so the top predictor is effectively an alternative expression of the label. Second, Algorithm 1 performs RFECV and SHAP feature selection before the train/test split, and only ADASYN and normalization are described as fold-confined. This allows the selected 20-feature set to be informed by test labels, inflating the reported test-set metrics. These are not mere omissions: the quoted Algorithm order and the quoted outcome definition exhibit the reduction directly. The model fitting itself is empirical and the reported point estimates are internally reproducible in principle, so the paper is not wholly circular, but the headline performance and the leading risk-factor interpretation are weakened by construction. The stated limitations acknowledge retrospective design and lack of external validation, but they do not address the label-feature code overlap or the pre-split feature selection. Score 6 reflects partial circularity in the central predictive claim rather than a fully derived or purely self-referential result.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical ML study and introduces no new physical entities. Its central performance claim rests on multiple unexamined assumptions about outcome timing and code semantics, plus hand-set thresholds and tuning choices, and on a feature-selection procedure that uses the full dataset before evaluation.

free parameters (6)
  • Missingness thresholds (30% rows, 20% columns) = 30% and 20%
    Records and features above these missingness rates were dropped; the thresholds are arbitrary and affect cohort size and feature set.
  • Feature subset size k=20 (RFECV/SHAP intersection) = 20
    The number of features retained was chosen by intersecting RFECV and SHAP top lists; k=20 is a hand-set target and affects all downstream metrics.
  • Age cutoff (55 years) = >55
    The elderly definition is a modeling choice that changes cohort composition and prevalence.
  • ICU stay cutoff (48 hours) = >=48h
    Admissions shorter than 48h were excluded; this is a cohort-defining choice.
  • CatBoost hyperparameters (learning rate, depth, regularization) = not reported
    Tuned by 5-fold grid search to maximize mean AUROC; final values are not reported, so replication requires re-tuning.
  • ADASYN sampling parameters (e.g., n_neighbors) = not reported
    Synthetic oversampling parameters are not stated; they change the training distribution and class balance.
assumptions (5)
  • domain assumption ICD-9 diagnosis codes 430-436 identify incident in-hospital stroke events
    The outcome label relies on these codes, but no timing information ties them to a new stroke after surgery; the paper does not validate the timing in this cohort.
  • domain assumption Clinical features from the first 24 hours of ICU stay are free of post-stroke physiology
    If the stroke occurred during or before that window, the features reflect the outcome, not predictors. This is not examined.
  • ad hoc to paper The 'cerebrovascular disease' comorbidity flag is distinct from the stroke outcome
    Both are derived from overlapping ICD-9 code ranges; the paper treats them as independent without demonstrating code-level separation.
  • domain assumption SHAP value directions correspond to true risk-factor relationships
    The SHAP narrative contradicts raw group differences in Table 3 for creatinine and heart rate, showing the assumption is unsafe.
  • ad hoc to paper Feature selection on the full dataset does not leak validation information
    RFECV/SHAP selection was applied before splitting, which violates standard leakage-free evaluation unless nested in CV.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Postoperative Stroke in Elderly SICU Patients: An Interpretable Machine Learning Model Using MIMIC Data." pith.science (2026). https://pith.science/paper/WLJWRURE

@misc{pith2026250603209,
  author       = {Pith},
  title        = {Pith review of: Predicting Postoperative Stroke in Elderly SICU Patients: An Interpretable Machine Learning Model Using MIMIC Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WLJWRURE}},
  note         = {Machine review of arXiv:2506.03209}
}
read the original abstract

Postoperative stroke remains a critical complication in elderly surgical intensive care unit (SICU) patients, contributing to prolonged hospitalization, elevated healthcare costs, and increased mortality. Accurate early risk stratification is essential to enable timely intervention and improve clinical outcomes. We constructed a combined cohort of 19,085 elderly SICU admissions from the MIMIC-III and MIMIC-IV databases and developed an interpretable machine learning (ML) framework to predict in-hospital stroke using clinical data from the first 24 hours of Intensive Care Unit (ICU) stay. The preprocessing pipeline included removal of high-missingness features, iterative Singular Value Decomposition (SVD) imputation, z-score normalization, one-hot encoding, and class imbalance correction via the Adaptive Synthetic Sampling (ADASYN) algorithm. A two-stage feature selection process-combining Recursive Feature Elimination with Cross-Validation (RFECV) and SHapley Additive exPlanations (SHAP)-reduced the initial 80 variables to 20 clinically informative predictors. Among eight ML models evaluated, CatBoost achieved the best performance with an AUROC of 0.8868 (95% CI: 0.8802--0.8937). SHAP analysis and ablation studies identified prior cerebrovascular disease, serum creatinine, and systolic blood pressure as the most influential risk factors. Our results highlight the potential of interpretable ML approaches to support early detection of postoperative stroke and inform decision-making in perioperative critical care.

Figures

Figures reproduced from arXiv: 2506.03209 by the authors.

Figure 1
Figure 1. Patient Selection Process Flowchart Based on MIMIC-III and [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Data Preprocessing and Feature Selection Pipeline Flowchart [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Model Training and Evaluation Workflow Flowchart Detailing [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: ROC curve for the training cohort September 23, 2025 17/27 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: ROC curve for the test cohort [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: SHAP Summary Plot Showing Feature Contributions and [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Ablation study of the CatBoost model. Each box represents the distribution of AUROC scores after removing one individual feature. The dashed red line denotes the baseline performance when all features are included [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 36 canonical work pages

  1. [1]

    Pathophysiology and Treatment of Stroke: Present Status and Future Perspectives

    Kuriakose D, Xiao Z. Pathophysiology and Treatment of Stroke: Present Status and Future Perspectives. International Journal of Molecular Sciences. 2020;21:7609. doi:10.3390/ijms21207609

    work page doi:10.3390/ijms21207609 2020

  2. [2]

    Influence of age and health behaviors on stroke risk: Lessons from longitudinal studies

    Kelly-Hayes M. Influence of age and health behaviors on stroke risk: Lessons from longitudinal studies. Journal of the American Geriatrics Society. 2010;58(Suppl 2):S325–S328. doi:10.1111/j.1532-5415.2010.03075.x

    work page arXiv 2010

  3. [3]

    Donnan GA, Fisher M, Macleod M, Davis SM. Stroke. The Lancet. 2008;371(9624):1612–1623. doi:10.1016/S0140-6736(08)60694-7

    work page doi:10.1016/s0140-6736(08)60694-7 2008

  4. [4]

    Stroke, Cerebrovascular accident; n.d

    World Health Organization. Stroke, Cerebrovascular accident; n.d. https://www.emro.who.int/health-topics/ stroke-cerebrovascular-accident/index.html

    work page

  5. [5]

    A contemporary and comprehensive analysis of the costs of stroke in the United States

    Girotra T, Lekoubou A, Bishu KG, Ovbiagele B. A contemporary and comprehensive analysis of the costs of stroke in the United States. Journal of the Neurological Sciences. 2020;410:116643. doi:10.1016/j.jns.2019.116643

    work page arXiv 2020

  6. [6]

    Economic burden of stroke across Europe: A population-based cost analysis

    Luengo-Fernandez R, Violato M, Candio P, Leal J. Economic burden of stroke across Europe: A population-based cost analysis. European Stroke Journal. 2020;5(1):17–25. doi:10.1177/2396987319883160

    work page doi:10.1177/2396987319883160 2020

  7. [7]

    Global, regional, and national burden of stroke, 1990-2016: A systematic analysis for the Global Burden of Disease Study 2016

    GBD 2016 Stroke Collaborators. Global, regional, and national burden of stroke, 1990-2016: A systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurology. 2019;18(5):439–458. doi:10.1016/S1474-4422(19)30034-1

    work page doi:10.1016/s1474-4422(19)30034-1 2016

  8. [8]

    Stroke risk factors, genetics, and prevention

    Boehme AK, Esenwa C, Elkind MSV. Stroke risk factors, genetics, and prevention. Circulation research. 2017;120(3):472–495. doi:10.1161/CIRCRESAHA.116.308398

    work page doi:10.1161/circresaha.116.308398 2017

Show all 39 references
  1. [9]

    Aging and ischemic stroke

    Yousufuddin M, Young N. Aging and ischemic stroke. Aging (Albany NY). 2019;11(9):2542–2544. doi:10.18632/aging.101931

    2019 doi

  2. [10]

    Stroke epidemiology: a review of population-based studies of incidence, prevalence, and case-fatality in the late 20th century

    Feigin VL, Lawes CM, Bennett DA, Anderson CS. Stroke epidemiology: a review of population-based studies of incidence, prevalence, and case-fatality in the late 20th century. The Lancet Neurology. 2003;2(1):43–53. doi:10.1016/s1474-4422(03)00266-7

    2003 doi

  3. [11]

    Human and economic burden of stroke

    Di Carlo A. Human and economic burden of stroke. Age and Ageing. 2009;38(1):4–5. doi:10.1093/ageing/afn282

    2009 doi

  4. [12]

    Outcomes in perioperative care

    Mantz J, Dahmani S, Paugam-Burtz C. Outcomes in perioperative care. Current Opinion in Anaesthesiology. 2010;23(2):201–208. doi:10.1097/ACO.0b013e328336aeef. September 23, 2025 24/27

    2010 doi

  5. [13]

    Impact of perioperative acute ischemic stroke on the outcomes of noncardiac and nonvascular surgery: a single centre prospective study

    Biteker M, Kayatas K, Türkmen FM, Mısırlı CH. Impact of perioperative acute ischemic stroke on the outcomes of noncardiac and nonvascular surgery: a single centre prospective study. Canadian Journal of Surgery. 2014;57:E55–E61. doi:10.1503/cjs.003913

    2014 doi

  6. [14]

    Risk Factors and Stroke Characteristic in Patients with Postoperative Strokes

    Dong Y, Cao W, Cheng X, Fang K, Zhang X, Gu Y, et al. Risk Factors and Stroke Characteristic in Patients with Postoperative Strokes. Journal of Stroke and Cerebrovascular Diseases. 2017;26(7):1635–1640. doi:10.1016/j.jstrokecerebrovasdis.2016.12.017

    2017 doi

  7. [15]

    Prevention of Stroke in Patients With Atrial Fibrillation: Anticoagulation Strategies and Beyond

    Kamel H, Navi BB, Sriram N, Hovsepian DM, Devereux PJ, Bove AA, et al. Prevention of Stroke in Patients With Atrial Fibrillation: Anticoagulation Strategies and Beyond. Stroke. 2018;49(4):1073–1080. doi:10.1161/STROKEAHA.117.018927

    2018 doi

  8. [16]

    Stroke prevention: from primary prevention to early rehabilitation

    Catanese L, Bogousslavsky J, Caso V. Stroke prevention: from primary prevention to early rehabilitation. Swiss Medical Weekly. 2017;147:w14447. doi:10.4414/smw.2017.14447

    2017

  9. [17]

    Stroke declines from third to fourth leading cause of death in the United States: historical perspective and challenges ahead

    Towfighi A, Saver JL. Stroke declines from third to fourth leading cause of death in the United States: historical perspective and challenges ahead. Stroke. 2011;42(8):2351–2355. doi:10.1161/STROKEAHA.110.610777

    2011 doi

  10. [18]

    Global Aging and our Futures

    Grinin L, Grinin A, Korotayev A. Global Aging and our Futures. World Futures. 2023; p. 1–21. doi:10.1080/02604027.2023.2204791

    2023

  11. [19]

    Predictors of early and late stroke following cardiac surgery

    Whitlock R, Healey JS, Connolly SJ, Wang J, Danter MR, Tu JV, et al. Predictors of early and late stroke following cardiac surgery. CMAJ. 2014;186(12):905–911. doi:10.1503/cmaj.131214

    2014 doi

  12. [20]

    Usefulness of the CHA2DS2VASc score to predict postoperative stroke in patients having cardiac surgery independent of atrial fibrillation

    Peguero JG, Issa O, Podesta C, Elmahdy HM, Santana O, Lamas GA. Usefulness of the CHA2DS2VASc score to predict postoperative stroke in patients having cardiac surgery independent of atrial fibrillation. The American Journal of Cardiology. 2015;115(6):758–762. doi:10.1016/j.amj...

    2015 doi

  13. [21]

    Predicting mortality in Sepsis-Associated acute respiratory distress syndrome: A machine learning approach using the MIMIC-III database

    Mu S, Yan D, Tang J, et al. Predicting mortality in Sepsis-Associated acute respiratory distress syndrome: A machine learning approach using the MIMIC-III database. Journal of Intensive Care Medicine. 2025;40(3):294–302

    2025

  14. [22]

    Predicting ICU Readmission in Acute Pancreatitis Patients Using a Machine Learning-Based Model with Enhanced Clinical Interpretability

    Chen S, Fan J, Si Y, Sun L, Alaei K, Pishgar E, et al. Predicting ICU Readmission in Acute Pancreatitis Patients Using a Machine Learning-Based Model with Enhanced Clinical Interpretability. medRxiv. 2025; p. 2025–05

    2025

  15. [23]

    Zhao QY, Liu LP, Luo JC, et al. A machine-learning approach for dynamic prediction of sepsis-induced coagulopathy in critically ill patients with sepsis: an integrated analysis of the MIMIC-IV and eICU-CRD databases. Research Square. 2020;doi:10.21203/rs.3.rs-125438/v1

    2020 doi

  16. [24]

    Retrospective Machine Learning Approach for Forecasting In-Hospital Death in ICU Patients After Cardiac Arrest

    Si Y, Sun L, Chen S, Fan J, Pishgar E, Alaei K, et al. Retrospective Machine Learning Approach for Forecasting In-Hospital Death in ICU Patients After Cardiac Arrest. medRxiv. 2025; p. 2025–05

    2025

  17. [25]

    MIMIC-III, a freely accessible critical care database

    Johnson AE, Pollard TJ, Shen L, Lehman LWH, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Scientific data. 2016;3:160035. doi:10.1038/sdata.2016.35. September 23, 2025 25/27

    2016 doi

  18. [26]

    MIMIC-IV, a freely accessible electronic health record dataset

    Johnson AE, Stone DJ, Celi LA, Pollard TJ. MIMIC-IV, a freely accessible electronic health record dataset. Scientific data. 2021;8(1):1–11. doi:10.1038/s41597-021-01001-2

    2021 doi

  19. [27]

    Validity of International Classification of Disease Codes to Identify Ischemic Stroke and Intracranial Hemorrhage Among Individuals With Associated Diagnosis of Atrial Fibrillation

    Thigpen J, Dillon C, Forster K, Henault L, Quinn E, Tripodis Y, et al. Validity of International Classification of Disease Codes to Identify Ischemic Stroke and Intracranial Hemorrhage Among Individuals With Associated Diagnosis of Atrial Fibrillation. Circulation: Cardiovascu...

    2015 doi

  20. [28]

    Machine Learning Prediction Models for Postoperative Stroke in Elderly Patients: Analyses of the MIMIC Database

    Zhang X, Fei N, Zhang X, Wang Q, Fang Z. Machine Learning Prediction Models for Postoperative Stroke in Elderly Patients: Analyses of the MIMIC Database. Frontiers in Aging Neuroscience. 2022;14. doi:10.3389/fnagi.2022.897611

    2022

  21. [29]

    Normalization methods for clinical data in machine learning: A systematic review

    Petersen A, Nielsen T, Jensen L. Normalization methods for clinical data in machine learning: A systematic review. Journal of Biomedical Informatics. 2017;75:128–137

    2017

  22. [30]

    The importance of the normality assumption in large public health data sets

    Lumley T, Diehr P, Emerson S, Chen L. The importance of the normality assumption in large public health data sets. Annual Review of Public Health. 2002;23:151–169

    2002

  23. [31]

    Chi-square Test and its Application in Hypothesis Testing

    Rana R, Singhal R. Chi-square Test and its Application in Hypothesis Testing. Journal of the Practice of Cardiovascular Sciences. 2015;1(1):69–71. doi:10.4103/2395-5414.157577

    2015

  24. [32]

    Machine Learning–Based Model for Prediction of Outcomes in Acute Stroke

    Heo J, Yoon JG, Park H, Kim YD, Nam HS, Heo JH. Machine Learning–Based Model for Prediction of Outcomes in Acute Stroke. Stroke. 2019;50(5):1263–1265. doi:10.1161/STROKEAHA.118.024293

    2019 doi

  25. [33]

    Machine Learning for Brain Stroke: A Review

    Sirsat MS, Fermé E, Câmara J. Machine Learning for Brain Stroke: A Review. J Stroke Cerebrovasc Dis. 2020;29(10):105162. doi:10.1016/j.jstrokecerebrovasdis.2020.105162

    2020

  26. [34]

    Stroke Prediction with Machine Learning Methods among Older Chinese

    Wu Y, Fang Y. Stroke Prediction with Machine Learning Methods among Older Chinese. Int J Environ Res Public Health. 2020;17(6):1828. doi:10.3390/ijerph17061828

    2020 doi

  27. [35]

    Perioperative Low Arterial Oxygenation Is Associated With Increased Stroke Risk in Cardiac Surgery

    Dunham AM, Grega MA, Brown CH, McKhann GM, Baumgartner WA, Gottesman RF. Perioperative Low Arterial Oxygenation Is Associated With Increased Stroke Risk in Cardiac Surgery. Anesthesia & Analgesia. 2017;125(1):38–43. doi:10.1213/ANE.0000000000002157

    2017 doi

  28. [36]

    A Machine Learning Prediction Model of Respiratory Failure Within 48 Hours of Patient Admission for COVID-19: Model Development and Validation

    Bolourani S, Brenner M, Wang P, McGinn T, Hirsch JS, Barnaby D, et al. A Machine Learning Prediction Model of Respiratory Failure Within 48 Hours of Patient Admission for COVID-19: Model Development and Validation. J Med Internet Res. 2021;23(2):e24246. doi:10.2196/24246

    2021 doi

  29. [37]

    A unified approach to interpreting model predictions

    Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems. 2017;30:4765–4774

    2017

  30. [38]

    ADASYN: Adaptive synthetic sampling approach for imbalanced learning

    He H, Bai Y, Garcia EA, Li S. ADASYN: Adaptive synthetic sampling approach for imbalanced learning. In: 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence). IEEE; 2008. p. 1322–1328. September 23, 2025 26/27

    2008

  31. [39]

    Development and validation of a prediction model for strokes after coronary artery bypass grafting

    Charlesworth DC, Likosky DS, Marrin CA, Maloney CT, Quinton HB, Morton JR, et al. Development and validation of a prediction model for strokes after coronary artery bypass grafting. The Annals of Thoracic Surgery. 2003;76(2):436–443. doi:10.1016/s0003-4975(03)00528-9. Septembe...

    2003 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.