Study design and data source

A retrospective analysis was conducted on data from 2,795 adult patients admitted for neurological/neurosurgical reasons between November 2022 and August 2024. Patients’ diagnosis was categorized and analyzed based on the International Classification of Diseases and Related Health Problems (ICD), Tenth Revision. The patients were categorized into neoplasm, vascular and functional groups. The TRIPOD Checklist: Prediction Model Development was followed as a reporting guidance in this study.

Data were extracted in October 2025 from the national Electronic Health Solutions (EHS) database, managed anonymously by Health Data Analytics (HDA) and provided in separate parts for the authors with unidentifiable patient IDs. Subsequently, the authors had no access to patients’ records.

Variables and data preparation

The unplanned readmissions were not distinguished from planned readmissions at EHS system. They provided us with the admission dates and causes of admissions. We tested all the admissions and then categorized them into planned and non-planned according to the unplanned readmission definition. The period calculated from the readmission date and the index discharge date; all those who were readmitted in a period of less than 30 days for floor index admissions and less than 90 days for critical care index admissions were candidates for unplanned readmission group. Then, the causes for all candidate readmission were tested; if these readmissions had any attribute that related to the index cause categorized as unplanned readmissions, all other admissions categorized as planned readmissions. Because of fragmented and inconsistent data provided by EHS system, we used only the data for the last two admissions for those patients who had multiple readmissions.

Data included the following independent variables: 1) demographics, 2) primary diagnosis (categorized into Neoplasm, Vascular, and Functional groups), 3) comorbidities; the EHS system had used the ICD coding and classification system to categorize all the comorbidities. The ICD Ninth Revision was used till August 2017 then the system updated into the ICD Tenth Revision. The EHS system provided all documented comorbidities for all patient involved in the study over a lookback period started from February 2015 till August 2024. All those comorbidities involved in the Charlson Comorbidity Index (CCI) calculation were involved in the analysis in addition to the calculated CCI scores, 4) geographic location (governorate); divided into northern, southern, and middle governorate. However, the CHAID model manipulated and merged governorate categories and ended with two subgroups that share same characteristics in relation to the dependent variable; the northern one versus the southern/middle one, 5) length of stay (LOS); which is the period from the admission and discharge dates for the same hospitalization, 6) admission site (critical care/general ward), and 7) key laboratory values; we used only the first result for each laboratory test at the index admissions. Data cleaning, merging, and transformation were performed using the Statistical Package for the Social Sciences (IBM SPSS) Statistics 27. We represent in Fig 1 the data extraction, processing, and analysis procedures.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Data extraction, processing, and analysis procedures.

https://doi.org/10.1371/journal.pone.0360077.g001

Data processing and cleaning

We analyzed the data for noise, missing values, and inconsistency using outlier detection and frequency description in the IBM-SPSS Program (version 27.0). We excluded from the analysis all duplicated redundant data, and all variables that were either not provided by the EHS database or provided with huge missing or/and illogical readings. As an example of excluded variables, height, weight, symptom data, physician and nursing notes, and certain laboratory tests (such as lactate dehydrogenase, high-density lipoprotein (HDL), albumin, cholesterol, and fasting blood sugar data).

Missing data assessment

Of the original 54 candidate variables, 12 were excluded a priori due to >20% missingness: height (41%), weight (38%), lactate dehydrogenase (52%), HDL (44%), albumin (35%), cholesterol (47%), fasting blood sugar (39%), and five clinical note fields (65 89%). For the remaining 42 variables, missingness was < 3% per variable; thus, complete case analysis was applied (final n = 2,795). No imputation was performed because missingness in retained variables was very low, and excluded variables had excessively high missing rates where imputation would be unreliable. This exclusion may introduce bias if missingness is not completely at random, which we now note as a limitation.

Predictive modeling and selection

The SPSS Modeler 18 was used to build and compare nine prediction models (e.g., CHAID, C5.0, Logistic Regression, Neural Networks). Model performance was evaluated based on overall accuracy and the Area Under the Curve (AUC). The CHAID model was selected as the final model due to its superior balance of accuracy (91.1%) and interpretability (AUC = 0.671). The CHAID is a decision tree algorithm that uses chi-square tests to identify the most significant predictor for splitting the data at each node. It handles categorical variables effectively and produces an intuitive tree structure. Hyperparameters were set to a maximum tree depth of three levels, with a statistical significance (p < 0.05) threshold for splitting. A train-test split method was used for validation.

Model development and validation

The dataset was randomly split into training (70%, n = 1,956) and test (30%, n = 839) sets using SPSS Modeler’s random number generator with a fixed seed (12345) to ensure reproducibility. No stratification by outcome was applied; the split preserved the overall readmission rate (training: 8.8%, test: 9.0%). Internal validation was performed on the held-out test set only; no cross validation was used. All preprocessing parameters (e.g., one hot encoding categories, z score means/SDs) were derived from the training set only and then applied to the test set to prevent data leakage.

Preprocessing and missing data

Categorical variables (sex, region, admission site, diagnosis group) were one hot encoded. Continuous variables (age, LOS, laboratory values) were standardized (z score) only for Neural Net 1 and Discriminant 1; tree-based models and logistic regression used raw values. Variables with >20% missing were excluded a priori (height, weight, lactate dehydrogenase, HDL, albumin, cholesterol, fasting blood sugar). For retained variables, complete case analysis was used (missingness <3%).

Hyperparameter settings (all nine models)

All nine models were trained using default IBM SPSS Modeler 18 parameters except where noted. CHAID used a maximum depth of 3, minimum parent size of 100, minimum child size of 50, α split = 0.05, α merge = 0.05, and chi square splitting. C5.1 used pruning severity of 75% and minimum cases per child = 2. Quest used α split = 0.05, pruning enabled, and minimum parent size = 50. C&R Tree 1 used max depth = 5, min parent size = 50, and min child size = 25. Neural Net 1 used an automatic one hidden layer, max training time of 15 minutes, and tolerance = 0.0001. Logistic Regression 1 used stepwise entry (p ≤ 0.05) and removal (p ≥ 0.10). Bayesian Network 1 used hill climbing, max parents = 3, and BIC scoring. Discriminant 1 used stepwise Wilks’ lambda with entry F = 3.84 and removal F = 2.71. Decision List used confidence = 0.9, min support = 2%, and max rules = 50. No hyperparameter optimization (grid search / cross validation) was performed.

Model comparison

Nine models were compared on the test set (S1 Table). The CHAID was selected because it demonstrated the highest discriminative ability among the interpretable tree-based models (AUC = 0.671) while maintaining a transparent decision-tree structure. Although several models achieved the same overall accuracy (91.1%), three models (C5.1, Quest, and C&R Tree 1) had AUC = 0.5 (no better than random), confirming that accuracy alone was uninformative in this imbalanced dataset.

Ethical considerations

Ethical approval was obtained from the Jordanian Royal Medical Services (Document No. T F-3/1–6626, Meeting No. 9/2024). Patient consent was waived as the analysis used fully anonymized data; this eliminates some risks for selection bias.

Sample characteristics and readmission rates

Among 2,795 patients, 249 (8.9%) experienced an unplanned readmission. As presented in S2 Table, among those who had floor index admissions; a total of 211 (8.5%) patients experienced 30-day unplanned readmissions, while among those who had critical care index admissions; a total of 38 (11.8%) patients experienced 90-day unplanned readmissions. The neoplasm group reported the higher incidence of unplanned readmissions (24%); the higher 30-day floor unplanned readmissions (22.7%) and 90-day critical care admissions (50%). Among those who had index admission into critical care units, 11.4% in the vascular group and 5.2% in the functional group experienced 90-day critical care unplanned readmissions. Returning into critical care units among the unplanned readmissions differed between the study groups. In the neoplasm group, five patients (9.6%) out of the 52 patients who had unplanned readmissions were returned into critical care units. Meanwhile, 5 patients (4.6%) out of 109 patients in the vascular group and three patients (3.4%) out of the 88 patients in the functional groups had returned to critical care units in the unplanned readmissions.

S3 Table shows that the diagnosis-related complications (e.g., tumor progression, disc issues) were the most frequent cause of unplanned readmission overall (45.8%). Among the vascular group, which reported the higher critical care unplanned readmission, recurrent brain hemorrhage followed by cerebrovascular accident (CVA) reported the higher causes of unplanned readmissions (29.4% and 18.3%). Spinal disc disease reported the most common cause of unplanned readmissions (59.1%) among the functional group, which had the higher floor unplanned readmissions.

Key predictors from bivariate analysis

The overall mean age was 56.1 years (SD = 17.6), and 59.5% were male. As shown in S4 Table, the unplanned readmission rate was higher for index critical care admissions (11.8%) compared to general ward admissions (8.5%). We used independent sample T test for continuous variables, while the chi-square test was used for all the categorical variables. The analysis yielded significant bivariate predictors of unplanned readmission for the entire cohort included a CCI score ≥3 (p = 0.004), index admission to a critical care unit (p = 0.05), hyponatremia (p = 0.04), and history of cerebrovascular disease (p = 0.003), diabetes (p = 0.01), and renal disease (p = 0.04). Group-specific analyses found that in the vascular group, a high CCI (p = 0.01), history of cerebrovascular disease (p < 0.001), ischemic heart disease (p = 0.05), and critical care admission (p = 0.02) were significant. In the functional group, older age (p = 0.03), diabetes (p = 0.01), and abnormal lab values (white blood cells (WBC) [p = 0.03], platelets[p = 0.02]) were significant. Meanwhile, only the length of stay was significant in the neoplasm group (p = 0.03).

The CHAID prediction model

Although the bivariate analysis used a prespecified CCI threshold (≥3) for univariable comparison, the CHAID algorithm treated CCI as a predictor and automatically identified the split points that best separated patients according to readmission risk. Consequently, the final CHAID tree produced different CCI groupings (0, 1–4, 5–6, and 7), with the highest-risk subgroup emerging at CCI ≥ 5.

As presented in Fig 2, the CHAID decision tree identified CCI as the primary (root) predictor (Chi-square = 66.37, p < 0.001), splitting patients into four nodes. Due to class imbalance (8.9% readmission rate), we report additional metrics. Following selection of the final CHAID model using the independent train-test comparison, the final decision tree was applied descriptively to the complete cohort (N = 2,795) to characterize clinically interpretable patient subgroups. Consequently, the confusion matrix and the derived sensitivity, specificity, precision, F1 score, and balanced accuracy reported below were calculated from the complete cohort and are intended to summarize the subgroup-classification characteristics of the final tree rather than to provide independent out-of-sample validation. Independent model validation is represented by the held-out test-set AUC (0.671). The model revealed distinct pathways:

1. High CCI (Scores 5–7; n = 365): Patients, predominantly with vascular diagnoses (CVA, hemorrhage) as presented in S5 Table, formed the highest-risk groups (readmission rates: 15.3% [n = 51] for CCI 5–6, 38.7% [n = 12] for CCI 7). No further significant predictors were found for this high-risk cohort.
2. Medium CCI (Scores 1–4; n = 1,751): For this node, age was the secondary predictor.
   • Age < 40 (n = 166): Sex was a third predictor, with females having a higher readmission rate (30.4%; n = 17). Neoplasm was found among 37.3% of these young females.
   • Age 40–79 (n = 1,513): Among this aging subgroup, vascular problems were the dominant mainly CVA (32.1%; n = 506) and brain hemorrhage (28.5%; n = 452). Length of stay (LOS) was a third predictor among these patients with LOS > 13 days had a higher readmission rate (13.8%; n = 20).
   • Age ≥ 80 (n = 72): This group had a very low readmission rate (1.4%; n = 1); presented primarily with brain haemorrhage (68.1%; n = 49).
3. CCI = 0 (n = 679): Geographic region was the secondary predictor. The CHAID model divided the geographic region into two main subgroups; the northern one (n = 110) and the middle/south regions (n = 569). Patients from the northern governorate had a significantly higher readmission rate (10%; n = 11) compared to those from the middle/south (3.5%; n = 20). This group consisted almost entirely of functional disorder patients (89.4%; n = 607), mainly spinal disc disease (72.5%; n = 492).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Final CHAID decision tree generated from the complete study cohort (N = 2,795) for descriptive characterization of clinically interpretable patient subgroups.

Independent model validation was performed separately using the held-out test set (n = 839).

https://doi.org/10.1371/journal.pone.0360077.g002

Clinical utility and external validation needs

We acknowledge that an AUC of 0.671 represents poor-to-fair discriminative ability, limiting the model's suitability for individual level prediction while still supporting exploratory population level risk stratification. This performance is comparable to existing readmission prediction models in neurology, where reported AUCs typically range from 0.60 to 0.73. For example, Sander et al. (2020) reported AUCs of 0.62–0.68 for neurosurgical readmissions, and Hu et al. (2025) achieved an AUC of 0.73 in stroke patients with diabetes. Although the overall model comparison was performed using an independent train-test split, the confusion matrix, sensitivity, specificity, precision, F1 score, and balanced accuracy were calculated descriptively from the complete cohort after development of the final CHAID tree. Consequently, these statistics should not be interpreted as independent validation metrics and are likely to be optimistically biased. Future studies should evaluate all classification metrics using external validation cohorts or independent test datasets. Assessing generalizability, external validation in independent Jordanian healthcare sectors (e.g., university hospitals, private facilities) is necessary before clinical implementation. Future work should also evaluate the model's impact on discharge planning and readmission rates through prospective interventional studies.

Limitations related to model performance and methodology

Several limitations should be considered. First, the low incidence of unplanned readmissions (8.9%) created a class imbalanced dataset; overall accuracy (91.1%) overstates performance, whereas the corrected full-cohort balanced accuracy (0.63) provides a more realistic estimate of risk-stratification performance. Three comparator models achieved the same accuracy but had AUC = 0.5 (no better than chance), underscoring that accuracy alone is inappropriate for imbalanced outcomes. Second, we used a simple 70/30 train test split without cross validation or temporal validation. Third, hyperparameters were prespecified (or left at defaults) rather than optimized, which may have underestimated some models’ potential. Fourth, complete case analysis excluded <3% of records but could still introduce bias. Fifth, the random split was not stratified by outcome, though training and test readmission rates were nearly identical. Sixth, the poor-to-fair discriminative performance (AUC = 0.671) and absence of external validation mean this model is not suitable for clinical implementation at present; it should be used only for hypothesis generation and risk stratification in research or quality improvement contexts. Future work should employ k fold cross validation, repeated resampling, synthetic oversampling, cost sensitive learning, time based splitting, and hyperparameter tuning to improve model sensitivity and generalizability.

In conclusion, this study provides an interpretable framework for population-level risk stratification of neurological readmissions rather than a highly accurate individual prediction model. Although the CHAID model demonstrated poor-to-fair discrimination (AUC = 0.671; balanced accuracy = 0.63), it successfully identified clinically meaningful high-risk patient subgroups that may support quality-improvement initiatives after external validation. The resulting patient profiles offer a hypothesis generating blueprint for targeted quality improvement efforts (e.g., enhancing transitional care for high comorbidity vascular patients, supporting young neuro oncology patients, and addressing regional access gaps for spinal care). However, the model’s poor-to-fair discriminative performance and lack of external validation indicate that it should currently be regarded as a research and quality-improvement tool rather than a clinically deployable prediction model.