Enhancing clinical reliability in pressure injury prediction: A conformal prediction approach with machine learning models

1.1. Machine learning for pressure injury prediction

ML methods have become increasingly relevant in healthcare due to their ability to model complex, nonlinear relationships and support both diagnostic and prognostic tasks across diverse clinical applications.^22–24

Evidence across multiple domains, including mental health, ²⁵ multimodal precision health, ²⁶ oncology, ²⁷ and healthcare management, ²⁸ demonstrates the potential of ML to enhance decision-making.

For PI prediction, ML methods have been successfully applied using diverse data sources. Recent advances have shown promising results in early PI prediction, capturing complex relationships among clinical features, and have outperformed traditional risk assessment tools.^29–31 Studies have used electronic health records (EHRs) to model PI development in hospital units.^32–34 Other approaches leverage unstructured clinical notes through Natural Language Processing, ³⁵ or integrate more complex representations such as graphical models, Bayesian networks, or deep learning models analyzing imaging data.^36–38

These ML models often outperform traditional tools by better capturing interactions among clinical variables and providing early identification of high-risk patients. However, clinicians need confidence measures to make safer decisions for their patients, and most existing approaches provide only point predictions, while clinical data is inherently noisy and uncertain. For clinicians, reliable confidence measures are essential to support safe decision-making.

Following the TRIPOD reporting principles, ML models were developed using baseline clinical variables to evaluate uncertainty-aware prediction sets under different modeling configurations. The integration of CP provides calibrated uncertainty estimates, supporting methodological evaluation of model reliability and robustness under limited sample sizes and potential data variability. ³⁹

1.2. Conformal prediction

CP provides a principled framework for quantifying uncertainty around individual predictions. It generates prediction sets with formal coverage guarantees, making the output easy to interpret and grounded in statistical validity. ⁴⁰ CP can be applied to any underlying ML model without requiring assumptions about the data distribution or additional model parameterization beyond choosing a confidence level.

Applications of CP span multiple fields, including biometrics, facial recognition, biochemistry, finance, and healthcare. In medical contexts, CP has supported tasks such as disease progression prediction under uncertainty, ⁴¹ reliable genomic medicine, ⁴² depression prediction, ⁴³ lung cancer detection, ⁴⁴ coronary artery disease diagnosis using multi-objective methods, ⁴⁵ and data cleaning in biomedical sciences. ⁴⁶

The standard CP methodology involves first selecting a nonconformity measure, then training an ML model, and third, computing nonconformity scores for the calibration and test sets. These steps enable the generation of valid, efficient uncertainty-aware predictions. ³⁹

The clinical characteristics of PI development make this problem especially well-suited for uncertainty-aware modeling. PIs are rare, multifactorial, and highly sensitive to variations in mobility, hemodynamic stability, moisture exposure, device-related pressure, and the quality of preventive care. Small, clinically plausible changes in these factors can substantially modify a patient’s accurate risk profile. As a result, two patients may receive the same predicted probability while differing markedly in the reliability of that estimate. CP directly addresses this challenge by producing statistically calibrated prediction sets that indicate when the model is confident and when predictions should be interpreted with caution, for instance, during atypical patient presentations, incomplete records, or shifts in clinical practice. Studies in other high-stakes clinical domains demonstrate that CP can effectively flag uncertain or unreliable predictions before they propagate into clinical decision-making, thereby enhancing trustworthiness and reducing diagnostic risk. ⁴⁷ This property aligns closely with the realities of PI prevention, where false negatives may result in avoidable harm, and false positives increase the preventive workload for nursing teams already operating under high demand. By signaling the degree of uncertainty for each patient-level prediction, CP enables more transparent, risk-aware decision support, thereby strengthening the clinical utility of ML-based PI prediction systems.

2. 1. Study population

The study adopted an observational, cross-sectional design and used data from a tertiary-level healthcare institution in an urban area of the Metropolitan Region of Santiago, Chile. Data collection was conducted through a probabilistic random sampling procedure implemented by trained healthcare professionals across multiple hospital wards to ensure the representativeness of the inpatient population. The sample size was calculated at a 95% confidence level, with a 5% margin of error and an anticipated attrition rate of 20%, yielding an initial recruitment target of approximately 300 patients. Following data cleaning procedures, specifically the exclusion of atypical observations, incomplete records, and duplicated entries, the final analytical sample comprised 245 patients.

Data were collected using standardized forms, and information was extracted from patients’ electronic health records. The instruments were specifically designed to capture comprehensive information on both patient-specific characteristics and variables associated with the risk and development of PIs. The data collection framework was organized into four main domains: (i) sociodemographic characteristics, (ii) clinical parameters, (iii) laboratory and diagnostic test requests documented by attending physicians, and (iv) PI-related variables encompassing occurrence, stage, and anatomical location.

PIs were classified in accordance with the criteria established by the National Pressure Injury Advisory Panel (NPIAP). The assessment followed a dichotomous classification approach. Patients exhibiting no evidence of skin or tissue damage consistent with the NPIAP definition, including those with intact skin displaying only transient, blanchable erythema, were categorized as negative for PI. Conversely, patients were classified as positive when at least one lesion met the diagnostic criteria for a PI, irrespective of its severity or stage (Stages 1-4 or suspected deep tissue injury). Accordingly, the presence of a single lesion at any anatomical site was sufficient for a positive PI classification. ⁴⁸

2.2. Data cleansing and formatting

The data wrangling and preprocessing stage involved transforming unstructured nursing datasets into a format suitable for analysis. Relevant nursing documentation was extracted from the Hospital Information System (HIS) for each patient record. From these source data, key clinical concepts associated with the development of PIs were identified. This process was conducted in close collaboration with expert clinicians to ensure that text-based information was systematically converted into standardized and structured variables.

Since HIS records contain multiple entries per patient, the preprocessing stage considered a comprehensive aggregation and conflict-resolution strategy. For this purpose, time-window aggregation is used to group entries within clinically meaningful periods. For categorical data, conflicts in entries were resolved by selecting the most critical or informative values. For frequency-based features, values were aggregated to reflect the intensity or frequency of relevant events.

2.3. Feature selection

Feature selection was performed using the Chi-square test for categorical variables and the t-test for numerical features. The Chi-square test assesses whether the observed frequency distribution of two categorical variables deviates significantly from the expected distribution under the assumption of independence. The t-test identifies features with significant mean differences between outcome groups, supporting their inclusion in the modeling stage. This procedure ensured that only variables with meaningful statistical relationships to PI development were included in the training stage, thereby reducing dimensionality, improving model interpretability, and enhancing computational efficiency. ⁴⁹

2.4. Feature encoding

Since the dataset comprised categorical variables, a feature encoding step was required to transform qualitative attributes into numerical representations suitable for ML algorithms. Two complementary methods were applied: One-Hot Encoding (OHE) and Target Encoding (TE).

OHE is a method that converts categorical variables into a numeric format without imposing any order or magnitude on the categories. OHE creates K − 1 new binary features, each indicating whether the observation belongs to a specific category, while one is dropped to avoid mutual information sharing among features. This means that each category k ∈ K gets its own slot, and each observation activates only one of them.

Another process for handling categorical features is TE. To understand TE, let’s consider a feature space in the form x = {x₁, …, x _p } and a target variable y _i ∈ {0, 1} for binary classification. The TE method transforms a categorical variable x _j ∈ {c₁, …, c _k } into numerical values based on the conditional expectation of the target y. ⁵⁰ Its application is shown in (1).

T E ( c k ) = P ( y i = 1 | C i = c k ) = ∑ i = 1 n 1 ( C i = c k ) y i ∑ i = 1 n 1 C i = c k

(1)

As a result, each category is replaced by its empirical mean. To prevent overfitting, a regularization method is applied, as shown in (2).

T E regularized ( c k ) = n c k y ¯ c k + α y ¯ n c k + α

(2)

The key distinction between OHE and TE lies in their impact on dimensionality. Whereas OHE expands the feature space by generating one binary column per category, which may aggravate the curse of dimensionality. Conversely, TE encodes each category into a single continuous value, making it particularly suitable for high-cardinality categorical variables while retaining predictive information. ⁵⁰

For both methods, the encoding is conducted in an out-of-sample manner to avoid data leakage. Specifically, for each resampling iteration, encoding parameters were estimated using the training subset and subsequently applied to the corresponding testing subset, ensuring that no information from validation outcomes influenced feature construction. In the case of TE, category-specific statistics derived from the outcome variable were computed only from training observations, thereby preventing any observation from contributing to its own encoded value. In the case of OHE, it was fitted on the training data to define the set of categorical levels, and the resulting information was applied unchanged to the validation data. This fold-wise preprocessing strategy guarantees separation between training and validation information, yielding unbiased performance estimates and mitigating optimistic bias arising from data leakage. ⁵¹

2.5. Cross-validation

Cross-validation (CV) is a resampling technique used in ML and Statistics to evaluate a model’s performance, robustness, and generalization on unseen data. It helps prevent overfitting and ensures that the model’s evaluation is not dependent on randomness due to a specific setting. ⁵² The basic idea is to repeatedly split the dataset into training and validation sets and compute performance metrics for each split. The results are averaged to obtain a more stable estimation of the performance metrics.

In this paper, a k-fold CV scheme was adopted. The dataset was divided into k approximately equal folds while preserving the distribution of the target variable (stratified CV). For each iteration, k − 1 folds were used for training, and the remaining fold served as the validation set. This process was repeated k times, ensuring that each observation was used once for validation and k − 1 times for training. The averaged results across folds provided reliable estimates of the models’ predictive performance for the PI classification task, along with the results of the CP framework embedded within this approach.

2.6. Machine learning models

A total of seven ML models for binary classification were evaluated during the analysis. The models included tree-based algorithms: Decision Tree (DT), Random Forest (RF), and Extreme Gradient Boosting (XGB) for sequential learning; a logit model: Logistic Regression (LR); and margin-based models: Support Vector Machines for classification (SVC) with different kernel types, particularly SVC with a linear kernel, SVC with a polynomial kernel, and SVC with a Radial Basis Function (RBF). The selection of such models reflects different learning paradigms suitable for handling both linear and nonlinear patterns presented in clinical data.

Although ensemble learning can combine heterogeneous models, in this study, we focused on a tree-based approach, specifically on RF. This selection is mainly due to the complexity of heterogeneous learning, which arises from the combination of different paradigms within an optimization framework. Furthermore, mixing heterogeneous models also increases the risk of overfitting when the sample size is limited, since there is insufficient data to train multiple models and perform CV reliably. ⁵³

2.7. Hyperparameter optimization

To ensure optimal model performance and fair comparisons between the classifiers, a grid search hyperparameter optimization process was performed for each model. This process was embedded within the CV loop. The sets of hyperparameters were chosen according to standard best practices and systematically explored to minimize the risk of underfitting or overfitting. Table 1 shows the ranges of parameters explored for each classification algorithm. This optimization step is essential, as model performance is influenced not only by data quality but also by the appropriate specification of the model structure. This process enhances the robustness, generalization, and clinical utility of the predictive model. ⁵⁴

OPEN IN VIEWER

Table 1.

Set of hyperparameters chosen for the optimization process conducted in the cross-validation setting.

Model	Hyperparameter	Set of values
DT	Splitting criterion	{Gini, entropy, log}
	Splitter	{Best, random}
	Maximum depth	[1,20]
	Size of samples in leaf	[2,20]
	Size of samples to split	[1,20]
	Number of features	{Sq. root, log}
	Class weights	{Standard, balanced}
LR	Penalty	{l1, l2}
	Penalization size	[1,100]
	Class weights	{Standard, balanced}
RF	Trees	[1,100]
	Splitting criterion	{Gini, entropy, log}
	Maximum depth	[1,20]
	Size of samples to split	[2,20]
	Size of samples per leaf	[1,20]
	Number of features	{Sq. root, log}
	Class weights	{Standard, balanced}
XGB	Number of trees	[10,100]
	Maximum depth	[1,20]
	Learning rate	[0.01, 1.0]
	Subsample	[0.2, 1.0]
	Sample by tree	[0.1, 1.0]
	Class weights	{Standard, balanced}
SVC	Kernel	{Linear, Poly, RBF}
	Control parameter	[0.1, 10]
	Degree	{2,3,4}
	Gamma	[0.1, 10]
	Class weights	{Standard, balanced}

Hyperparameter ranges shown in Table 1 were selected to cover plausible model complexity in a clinical context, aiming to reduce the risk of overfitting in a modest sample and to capture nonlinear interactions. Furthermore, the ranges keep the search computationally feasible within the nested cross-validation procedure.

The hyperparameter search setting is an exhaustive search looking for all combinations of parameters within the defined set of potential values. The choice of the specific values was based on common practices in applied ML. For the DT model, given the sample size, we used a bounded, theory-consistent search space that spans from strongly regularized trees to moderately flexible trees. In the case of LR, the regularization is addressed with two methods ({l₁, l₂}), using a penalization parameter that ensures the search covers both heavily regularized and lightly regularized regimes. For RF and XGB, moderated complexity models were tested, enforcing strong regularization controls. The number of estimators (trees) was capped at 100 because RF performance typically stabilizes beyond moderate ensemble sizes, especially in small datasets. In the case of XGB, hyperparameters were selected to balance predictive flexibility with strong regularization, which is critical in small datasets. In this sense, tree depth, learning rate, and number of trees in the boosting rounds jointly control model complexity and overfitting issues. Finally, in the case of SVC, if kernels are too flexible, the models can easily overfit, so the grid explores different functional forms to control complexity through regularization parameters.

The class imbalance issue was addressed by assigning class weights proportional to each category’s frequency. This approach prevents the models from biasing towards the majority class and improves their ability to classify the minority class. ⁵⁵ The use of this approach, instead of artificial sampling methods, was primarily due to the limited sample size. In small datasets, synthetic samples (e.g., SMOTE, undersampling or oversampling methods) may increase the risk of overfitting by introducing unrealistic patterns that are not present in the actual data. ⁵⁶

2.8. Performance metrics for binary classification

The performance metrics for the classification task are based on the confusion matrix generated from the actual and predicted values. This tool provides a tabular summary that compares predicted and true class labels. Its structure is shown in Table 2.

OPEN IN VIEWER

Table 2.

Confusion matrix for a binary classification task.

​	Predicted positive	Predicted negative
Positive	True Positive (TP)	False Negative (FN)
Negative	False Positive (FP)	True Negative (TN)

From the Confusion Matrix, a set of metrics is computed to measure binary classification performance. Specifically, the Accuracy, Precision, and F1-Score are computed directly from the values provided by the classification models. Furthermore, the Area Under the Curve (AUC) is calculated from the Receiver Operating Characteristic (ROC) curve, which plots the True Positive Ratio (TPR) against the False Positive Ratio (FPR). ⁵⁷ Furthermore, the Brier-Score is computed to provide a more nuanced evaluation of probability, whereas standard accuracy and precision metrics only look at the final classification label. It is worth mentioning that Accuracy, Precision, F1-Score, and AUC metrics look for the highest possible value, whereas the Brier-score looks for the lowest possible value. ⁵⁸ Later, as each classification metric is computed across k folds, pairwise comparison methods are used to test for significant differences in model performance across the various metrics.

2.9. Statistical comparison of classification performance

The classification metrics are calculated across folds for various models. For this setting, Friedman’s test is employed, as performance metrics are paired within folds, providing evidence of significant differences in performance across models. Once the Friedman test is conducted and significant differences are identified, the Wilcoxon signed-rank post hoc test is used to determine which models exhibit significant differences in classification performance across folds. This test is suitable as it handles the paired structure in the data while controlling for the family-wise error rate. ⁵⁹

2.10. Conformal prediction Framework

To understand the basis of CP, let’s consider z = x × y the input-output space forming a dataset in the form z = z₁, z₂, …, z _n , such that z _i = (x _i , y _i ) is assumed to be exchangeable. This latest assumption is a weak form of the iid assumption for statistical models. Given a new input x_n+1, the goal is to construct a prediction set Γ _n (x_n+1) ⊆ y such that:

P ( y n + 1 ∈ Γ n ( x n + 1 ) ) ≥ 1 − α

(3)

α i = 1 − p ^ ( x i ) y

(4)

p ^ y = | { i : α i ≥ α n + 1 ( y ) } | + 1 m + 1

(5)

Γ n ( x n + 1 ) = { y : p y ≥ α }

(6)

This setting can be applied to any ML algorithm. Furthermore, no additional parameterization is required to select the non-conformity measure. As stated above, a key advantage of CP is the validity of predictions provided by its application, considering the assumptions of exchangeability and randomness.^60,61

2.11. Conformal prediction metrics

To contrast the ML metrics and provide an objective measure of the CP results, the coverage and efficiency metrics are employed. Let n be the number of test samples, let y _i be the true label for the i-th test point, and let Γ i α be the prediction interval at the confidence level 1 − α for an input x _i . The coverage is calculated as shown in (7).

C o v e r a g e = 1 n ∑ i = 1 n 1 ( y i ∈ Γ i α )

(7)

E f f i c i e n c y = 1 n ∑ i = 1 n | Γ i α |

(8)

It is worth mentioning that the metrics shown in (7) and (8), along with the CP framework, benefit from high-performing models. Particularly, the selection of the models with the best performance yields the following properties ⁴² .

• Narrower prediction sets: A high-performing model assigns higher scores to the correct class and lower scores to others, allowing CP to generate more informative sets.

The methodological framework was implemented in two independent stages to avoid information leakage between model optimization and uncertainty quantification. In the first stage, multiple ML models were trained and tuned using a k − fold cross-validation setting to identify the best-performing models and their optimal configurations. Model selection was based on predictive performance from out-of-fold validation. In the second stage, CP was applied as a post-hoc uncertainty quantification framework. The hyperparameters selected in the first stage were fixed, and the selected models were retrained within a new k − fold cross-validation procedure. Out-of-fold predictions generated in this stage were used to compute CP sets and evaluate the coverage and efficiency metrics shown in (7) and (8).

3.1. Occurrence of pressure injuries

The total incidence of PIs was approximately 27% among patients. About 20% of the injuries occurred in the Surgery ward, followed by the MQ IND and the critical patient wards, with 18% and 17%, respectively. About 62% of the patients were male. The average age of the patients was 59 years. Females were older, with an average age of 65 years, while males had an average age of 57 years. In terms of risk, about 39% of patients were categorized as high-risk, 22% as medium-risk, and 33% as low-risk, all during their admission to the health center units.

3.2. Univariate analysis of features for LPP prediction

The results of the feature selection for LPP prediction are shown in Table 3, along with the results of the Chi-Square test and the t-test. The first and second columns display the feature names and the categories, respectively. The third and fourth columns show descriptions for the group with PI and the group without PI. The fifth and sixth columns display the results of the feature selection method, including the statistic and p-value for each test. The age shows significant differences between the PI and non-PI groups. The Hospital Ward, the Risk assessment, the use of mattresses, the Containment, and the presence of Dermatitis are significant predictors of PI development.

OPEN IN VIEWER

Table 3.

Comparison of demographic, clinical, and care-related characteristics between patients with PI and without PI (NPI). Continuous variables are expressed as mean ± standard deviation, and categorical variables as counts. Statistics and p-values correspond to Chi-square test results for categorical data and t-test results for numerical variables.

Feature	PI	NPI	Statistic	p-value
Age	64 ± 20	58 ± 19	3.03	0.003*
Gender	​	​	0.00	∼ 1.00
Female	25	68	​	​
Male	41	111	​	​
Ward	​	​	19.68	0.006*
CI Huap	6	12	​	​
MQ IND	12	9	​	​
Medicine	9	47	​	​
Burned	6	11	​	​
TMT	6	19	​	​
UPC	11	19	​	​
Emergency	3	26	​	​
Surgery	13	36	​	​
Risk	​	​	27.81	∼ 0 . 0 *
High	41	54	​	​
Medium	16	39	​	​
Low	7	75	​	​
N.R.	2	11	​	​
Mattress	​	​	28.14	∼ 0 . 0 *
Yes	47	60	​	​
No	10	119	​	​
Position	​	​	6.19	0.102
Yes	23	42	​	​
No	42	133	​	​
DVA	​	​	0.28	0.595
Yes	3	4	​	​
No	63	175	​	​
Mobilization	​	​	0.71	0.398
Yes	5	7	​	​
No	61	172	​	​
Containment	​	​	8.63	0.003*
Yes	16	16	​	​
No	50	163	​	​
Dermatitis	​	​	5.35	0.021*
Yes	8	58	​	​
No	6	173	​	​

3.3. Model performance

Figure 2 compares how the choice of encoding techniques impacts model performance across different algorithms. The effect of encoding is model-dependent. Some algorithms benefit from TE, while others perform better with OHE. The DT model shows slightly higher variability with OHE across all metrics. TE yields more stable results, as highlighted by the AUC and F1-score metrics. The LR model shows similar results across encoding methods. Since this model is linear, OHE often works better, as it preserves feature independence. The RF model performs better with OHE across all metrics, particularly in precision and accuracy. The performance of the SVC (linear and polynomial) fluctuates, but TE tends to yield more consistent scores, especially in AUC and F1-score. The SVC RBF displays high variability with both encodings but shows slightly better median metrics with TE. The XGB consistently achieves the highest performance with TE, thanks to its compact numerical representations.OPEN IN VIEWER

Overall, the AUC and accuracy tend to be higher with TE for nonlinear models (SVCs, XGB). The RF with OHE shows outstanding precision. In terms of the F1-score, the models achieve a more balanced performance under TE, indicating a better tradeoff between precision and recall. In summary, TE offers advantages for complex nonlinear models, possibly by reducing dimensionality and introducing numerical features. Details of the improvements in the hyperparameter optimization process with cross-validation are shown in Table S11 in the Supplementary Material.

The results of the pairwise comparisons depend on the feature encoding methods used during fitting. The Accuracy shows significant differences with both encoding methods (p-values ≤ 5 % ). The Precision yields significant differences only with OHE (p-values ≤ 1 % ). The AUC shows significant results for both methods (p-values ≤ 10 % ). The F1-score yields significant differences only for the TE method (p-value ≤ 1 % ). Finally, the Brier-Score yields significant results for both methods (p-values ≤ 5 % ).

Pairwise comparisons using the Wilcoxon signed-rank test reveal significant differences, primarily for the RF and XGB models. The RF shows significantly better accuracy than DT, LR, SVC Linear, and SVC RBF, using both OHE and TE methods (p-values ≤ 10 % ). The XGB performs significantly better than LR and SVC Linear in terms of accuracy using both encoding methods (p-values ≤ 5 % ). Regarding precision, RF and XGB appear to be the best-performing models, with significant differences compared to all other models, especially with TE encoding (p-values ≤ 10 % ). In terms of the Brier-Score, the XGB model outperforms all models except RF at the 10% significance level, while RF performs significantly better than DT and LR depending on the encoding method. The different metrics show significant differences in particular encodings. The RF and XGB models perform slightly better than SVC Linear, SVC Poly, and SVC RBF, using OHE and TE, in terms of AUC. The F1-score shows no significant differences. Details on the results of the pairwise comparisons are shown in Tables S1 to S10 in the Supplementary Material.

In summary, across all evaluation metrics, the boxplots in Figure 2 indicate that the selection of the encoding method is model-dependent. Nevertheless, RF and XGB yield superior performance in the classification task, which is confirmed by the Friedman test (see Supplemental Material for tables with Friedman test results), showing significant performance variation across algorithms. The post-hoc Wilcoxon signed-rank test with correction further reveals that RF and XGB perform significantly better than the remaining models at the 10% significance level. Furthermore, both models yield no difference in performance, as shown in Table 4, which includes a summary of the metrics for binary classification for both models and the two encoding methods. Thus, both visual evidence and statistical tests consistently highlight the dominance of RF and XGB in this particular classification task. Moreover, the TE encoding shows strong predictive performance and better reliability, which is highlighted by the results of Precision with a performance about 6% better for RF and 10% better for XGB. Given this, TE encoding is selected for further analysis using the CP framework.

OPEN IN VIEWER

Table 4.

Summary with average of performance metrics for RF and XGB models per encoding method, values in parenthesis denote the standard deviation within the cross-validation framework.

Metric	Model	OHE (%)	TE (%)
Accuracy	RF	82.9 (5.12)	84.1 (4.65)
	XGB	82.4 (2.23)	84.5 (4.31)
AUC	RF	86.8 (5.69)	88.5 (4.23)
	XGB	85.4 (5.62)	85.5 (5.08)
Precision	RF	76.5 (12.3)	81.5 (15.9)
	XGB	70.7 (12.8)	80.0 (14.0)
F1-score	RF	65.1 (5.10)	66.3 (5.53)
	XGB	60.6 (5.53)	64.3 (6.64)
Brier-Score	RF	13.3 (1.91)	13.0 (2.24)
	XGB	13.3 (1.52)	12.6 (1.86)

3.4. Reliability of proposed models

Different values of confidence are tested, as shown in Figure 3. The chart above shows the ideal target coverage as a red dashed line. The closer the empirical line is to the red dashed target, the better the calibration of the conformal predictor. The blue (XGB) and green (RF) curves show the actual coverage achieved by both models. The models slightly underperform for some values of α. The chart below shows the average prediction set size for different values of α. As α increases, confidence decreases, and the prediction sets become smaller. In this sense, XGB tends to yield slightly smaller prediction sets than RF for the same α values, resulting in tighter intervals or fewer predicted labels on average, without compromising the model’s coverage.OPEN IN VIEWER

The choice of the α value depends on a trade-off between reliability and efficiency. As shown in Figure 4, the behavior of the CP metrics is model-dependent. In the context of PI, as pressure injury prevention has asymmetric costs, with missed high-risk cases leading to increased hospitalizations and serious harm, the CP framework must achieve high but practical coverage with actionable outputs. In this regard, if the α value is too small, prediction sets become large, and the model becomes uninformative; thus, the choice of α is analogous to selecting clinical risk cutoffs. Figure 4 shows the behavior of the coverage and set size for various α values. Regarding coverage, both models yield similar results; however, for values of α ≥ 0.10, the RF model shows greater variability than the XGB. In terms of efficiency, the RF model shows larger set sizes for α values above 10%. For this setting, a value of α = 0.10 ensures 90% coverage, representing a pragmatic compromise and reliability guarantees while maintaining prediction sets sufficiently specific to support actionable clinical decisions. Furthermore, this choice aligns with the goal of uncertainty-aware decision support rather than fully automated classification frameworks.OPEN IN VIEWER

The behavior of the empirical coverage being slightly higher than the nominal level is expected in CP frameworks due to the discrete nature of nonconformity scores and finite sample effects, which lead to conservative prediction sets. In practice, mild coverage is preferred to under-coverage for clinical decision-support contexts, as it ensures that the true outcome is included in the prediction set more frequently than the nominal guarantee.

Furthermore, at α = 0.10, XGB achieves a set size of 1.34. In a binary classification setting, this implies that the model produces a single, confident prediction in the majority of cases (around 66%), and assigns two-label prediction sets only when model uncertainty is high (around 34%).

Figure 5 shows a breakdown of CP setting results in terms of coverage and set size for an α = 10%. The CP prediction framework maintained coverage close to the nominal level across all patient risk groups. Low-risk patients exhibit the highest efficiency, with prediction sets that almost always contain a single label, indicating high confidence in model predictions. In contrast, medium- and high-risk patients produced larger prediction sets, reflecting increased uncertainty in the classification task. Across models, RF tends to produce larger prediction sets, whereas XGB exhibits more conservative behavior.OPEN IN VIEWER