Early Detection of Deep Tissue Pressure Injury in Intensive Care Using Hemodynamics‐Based Machine Learning: A Retrospective Cohort Study

ABSTRACT

This study examines factors associated with deep tissue pressure injury and develops interpretable machine learning models for early risk prediction in adult ICU patients. This retrospective observational cohort study included 336 adult intensive care unit patients, of whom 211 developed deep tissue pressure injury and 125 remained pressure injury‐free. For patients who developed deep tissue pressure injury, haemodynamic, laboratory and nursing variables from the 24 h before injury onset were analysed using a physiological time‐at‐risk framework. For pressure injury‐free patients, corresponding variables from the first 24 h after intensive care unit admission were used. Six supervised machine learning classifiers were developed and internally validated using cross‐validation and hyperparameter optimization. All models showed good predictive performance. Extreme gradient boosting achieved the highest discriminative ability, with an area under the receiver operating characteristic curve of 0.976. The most influential predictors across feature selection methods were low‐molecular‐weight heparin use, norepinephrine duration, lower blood pressure values, immobility, nutritional risk, antiplatelet therapy and chronic disease profile. Routinely documented nursing and haemodynamic indicators obtained within a clinically relevant 24‐h risk window can support accurate early risk stratification for deep tissue pressure injury in adult intensive care unit patients. A physiology‐informed and interpretable machine learning approach may improve recognition of patients at imminent risk.

Key Points

• DTPI risk may be detectable before visible skin changes using routinely collected nursing and hemodynamic data from the 24 h preceding DTPI onset.

• A physiology‐informed 24‐h time‐at‐risk framework enabled clinically relevant early prediction, with XGBoost showing the strongest discriminative performance (AUC 0.976).

• Perfusion‐related indicators, vasopressor exposure, immobility, and nutritional vulnerability provided clinically meaningful signals associated with subsequent DTPI development.

• These findings support DTPI as a cumulative and time‐sensitive physiological process and provide a foundation for nurse‐centered, explainable decision‐support tools for earlier prevention.

1. Introduction

Pressure injuries (PIs) remain a major and persistent challenge in intensive care units (ICUs), where immobility, haemodynamic instability, vasopressor therapy and multiple comorbidities substantially increase the risk of deep tissue damage and subsequent complications [1, 2]. Among all PI subtypes, deep tissue pressure injuries (DTPIs) represent a particularly severe and clinically elusive form. They originate at the muscle–bone interface, where sustained mechanical deformation can initiate primary, direct deformation damage, while secondary inflammatory‐oedema related damage and tertiary ischaemic damage interact with impaired perfusion and individual tissue tolerance [1]. These mechanisms may act cumulatively and need not occur in a strictly linear sequence, consistent with the vicious cycle of cell and tissue damage described by Gefen et al. [1]. Because this subclinical deterioration progresses rapidly, DTPIs are frequently recognised late, leading to delayed intervention, poor prognosis and significant declines in patient quality of life [3, 4]. DTPI prevalence in ICU populations is highly variable but consistently elevated, reflecting the interplay of immobility, circulatory insufficiency, and sedation‐related sensory loss. Recent studies in high‐risk ICUs reported rates of up to 30%–50% in these cohorts, underscoring both the vulnerability of this population and the inadequacy of current prevention and surveillance strategies [5, 6, 7]. This epidemiological burden highlights the urgent need for early detection strategies that can identify evolving deep tissue vulnerability and perfusion compromise before irreversible deep tissue damage occurs [1, 3, 4]. These physiological alterations support the concept of a proximal risk window preceding DTPI manifestation, during which evolving microvascular compromise and haemodynamic instability may provide clinically meaningful signals for early detection and targeted preventive interventions [1, 5, 8, 9].

Conventional risk assessment tools, such as the Braden and Norton scales, have limited predictive capability in critically ill patients because they fail to capture the dynamic physiological fluctuations and haemodynamic instability characteristic of ICU settings [10, 11]. Nurses frequently report that these tools overlook rapidly changing, individualised risk factors, thereby reducing their usefulness for early detection of DTPIs [12, 13]. Recent evidence suggests that DTPI development is strongly associated with hypotension, hypoperfusion, vasopressor exposure and prolonged mechanical ventilation, factors commonly encountered in critically ill patients but often overlooked in traditional risk‐scoring systems [2, 14]. The growing literature also indicates that biochemical markers such as low haemoglobin, hypoalbuminemia and elevated C‐reactive protein have been reported as indicators of impaired tissue perfusion and increased PI risk in prior studies [15]. Additionally, parameters including the frequency and duration of hypotensive episodes, lactate elevation, perfusion index and capillary refill time may serve as physiologically meaningful early indicators of DTPI development [16, 17]. In contemporary ICU environments, large‐scale, multidimensional streams of nurse‐documented vital signs, laboratory values and treatment‐related parameters are continuously generated. These routinely collected clinical data provide an opportunity to apply ML methods to support early risk stratification and time‐sensitive clinical decision‐making [18, 19].

ML approaches enable the analysis of complex nonlinear interactions within clinical datasets and have shown potential to outperform traditional risk assessment scales by generating individualised, dynamic risk estimates [11, 17]. Prior studies using algorithms such as Random Forest and Support Vector Machines have reported improved accuracy, sensitivity and specificity compared with conventional risk tools [9, 16, 20, 21]. However, most existing work has focused on general PI occurrence prediction and has rarely developed DTPI‐specific models incorporating haemodynamic time‐sensitive patterns and perfusion‐related physiological signals. Therefore, this study aimed to develop and validate an ML‐based model for early prediction of DTPI in ICU patients within a physiologically defined proximal risk window, with a specific focus on hypotension, hypoperfusion and vasopressor exposure. By identifying early haemodynamic predictors derived from nurse‐collected physiological and laboratory data, this research seeks to enhance early detection and prevention strategies and support timely, evidence‐based clinical decision‐making in critically ill patients.

2. Methods

2.1. Study Aim

This retrospective cohort study aimed to (1) examine associations between DTPI development and key physiological, laboratory and therapeutic parameters within a physiologically defined proximal risk window, including haemodynamic instability (hypotension, impaired perfusion), metabolic indicators (haemoglobin, albumin) and vasopressor exposure in adult ICU patients; and (2) develop and validate ML–based predictive models for early DTPI risk stratification.

2.2. Research Questions

1. Which routinely documented haemodynamic, laboratory and therapeutic parameters are most strongly associated with DTPI development in critically ill adult ICU patients?

2. Do ML‐based models demonstrate improved predictive performance compared with conventional nursing risk assessment tools (Braden Scale, APACHE II and NRS‐2002) and which algorithm achieves the highest discriminatory accuracy?

2.3. Study Design and Setting

This retrospective observational cohort study with embedded machine learning model development and internal validation was conducted in the adult medical, surgical and general ICUs of Koç University Hospital, a tertiary academic medical centre in Istanbul, Türkiye, with data analysis performed between October 15, 2024 and October 1, 2025. Retrospective clinical data were extracted from routinely documented nursing records, physiological monitoring data, laboratory reports and treatment charts for all eligible ICU admissions occurring between January 1, 2022 and April 30, 2025. The study followed the STROBE guidelines for observational research [22] and the TRIPOD‐AI recommendations for transparent reporting of ML‐based prediction model development and validation [23].

The temporal design of the study was also informed by contemporary mechanobiological concepts of pressure injury development, particularly the vicious cycle model described by Gefen et al. [1], which conceptualises deep tissue damage as a time‐dependent process triggered by sustained mechanical deformation and involving primary, direct deformation damage, secondary inflammatory‐oedema related damage and tertiary ischaemic damage, with the timing and severity of these processes influenced by individual tissue vulnerability. Additionally, this period was chosen to minimise potential confounding factors related to the COVID‐19 pandemic, which substantially impacted ICU admission patterns and clinical workflows during earlier phases. The study period was selected to ensure a sufficiently large and clinically heterogeneous ICU cohort under stable operational conditions, thereby reducing temporal variability and enabling the inclusion of diverse haemodynamic and physiological profiles necessary for model development and internal validation.

2.4. Participants and Sampling Criteria

Eligible participants were adult patients (≥ 18 years) admitted to the ICU with an ICU length of stay of ≥ 24 h, thereby ensuring adequate exposure to ICU‐specific risk factors for DTPI. All ICU admissions during the study period were screened consecutively for eligibility and patients who met the predefined inclusion criteria and had complete retrospective haemodynamic and laboratory documentation were enrolled. Accordingly, this sampling approach reflects a consecutive, eligibility‐based inclusion strategy consistent with retrospective cohort methodology [24]. Patients were divided into two groups based on their outcome: those who developed DTPI during their ICU stay and those who remained free of any PI during their ICU stay (comparative group).

Exclusion criteria consisted of

• pre‐existing Stage 1–4 PIs,

• mucosal membrane injuries,

• unstageable PIs identified at ICU admission and

• incomplete retrospective haemodynamic or laboratory data.

Fewer than 5% of cases were excluded due to missing or incomplete data, thereby minimising selection bias. A total of 336 patients met all inclusion criteria, comprising 211 who developed a DTPI and 125 who remained PI‐free throughout their ICU stay. This cohort represents all eligible ICU admissions identified through systematic daily surveillance of nursing documentation and PI records and provided an adequate dataset for both statistical analyses and the development and validation of ML models.

2.5. Outcome Definition

The primary outcome was the occurrence of a DTPI during the ICU stay, as defined according to the NPIAP/EPUAP/PPPIA (2019) international guidelines [25]. Daily skin assessments were routinely performed by bedside clinical nurses, who identified and staged PI as part of standard ICU care. Certified wound, ostomy and continence nurses (WOCNs) conducted independent skin assessments twice weekly. All suspected DTPIs identified by clinical nurses were subsequently reviewed and confirmed by the WOCNs and DTPIs identified during WOCN assessment days were documented directly by the WOCNs. For this study, all DTPI cases were retrospectively identified based on WOCN‐confirmed diagnoses. The WOCNs were blinded to haemodynamic and laboratory parameters to prevent bias in outcome classification. Any discrepancies were resolved through consensus. Only DTPIs newly developed after ICU admission were included in the primary analysis. The time of DTPI onset was defined as the first documented assessment in which DTPI diagnostic criteria were met. Patients in the DTPI group were considered at risk from ICU admission until DTPI onset, whereas patients in the PI‐absent group were considered at risk until ICU discharge, transfer or death. Other PI stages (Stage 1–4, unstageable or mucosal membrane PIs) were recorded for descriptive purposes only and were not included as part of the primary outcome. The secondary outcome comprised descriptive analyses of the time from hospital admission to first documented DTPI and its anatomical distribution. Patients classified as PI‐absent were those who remained free of any PI, including DTPI and other PI stages, throughout their ICU stay and hospitalisation. This classification was adopted to ensure a clearly defined comparator group for evaluating DTPI‐specific risk profiles and to minimise potential misclassification of outcomes related to progression between PI stages.

2.6. Data Sources and Variables

Data were collected retrospectively using a pre‐designed case report form (CRF) developed in accordance with NPIAP definitions and established ICU‐specific risk factors for PIs. All variables were defined a priori to ensure consistency and reproducibility for feature engineering and subsequent ML analyses. Clinical information was initially extracted manually from bedside flow sheets, nursing documentation and laboratory reports and was then entered into a secure, password‐protected electronic database. Data collection was conducted by a five‐member nursing team. Two nurses were nationally certified Wound, Ostomy and Continence Nurses (WOCN) and the remaining three were senior Registered Nurses with extensive experience in wound assessment and ICU care. For all suspected DTPI cases, diagnostic confirmation was performed exclusively by the WOCN‐certified nurses through an independent dual‐review process, followed by consensus resolution of discrepancies to standardise documentation and enhance inter‐rater reliability in accordance with international PI guidelines [25]. To maintain data integrity, the research team followed a predefined abstraction protocol, performed periodic quality checks and maintained an audit trail documenting all modifications made during data entry. Unclear or conflicting clinical entries were resolved through consensus discussions and, when necessary, the original bedside notes were re‐examined to verify the clinical context. All electronic entries were subsequently reviewed for completeness and accuracy by the full nursing team, with final verification led by the two WOCN‐certified nurses. Internal cross‐checking procedures were conducted by abstraction reviewers, whose verification was limited to the relevant clinical documentation and did not involve review of haemodynamic or laboratory data, thereby minimising potential bias.

2.7. Time‐At‐Risk Framework

To accurately capture physiological changes preceding DTPI onset, hemodynamic and laboratory variables were extracted using a deliberately defined time‐at‐risk framework. For patients who developed a DTPI, all haemodynamic and laboratory data were obtained from the 24 h immediately preceding DTPI identification. This interval was selected as a proximal risk window to preserve the DTPI‐specific physiological focus of the study rather than characterise general PI risk at ICU admission. The time‐at‐risk approach was informed by the contemporary mechanobiological vicious cycle framework described by Gefen et al. [1], in which sustained mechanical deformation acts as the triggering event and may lead to primary, direct deformation damage, secondary inflammatory‐oedema related damage and tertiary ischaemic damage that contribute cumulatively to cell and tissue injury. Accordingly, physiological conditions immediately preceding DTPI recognition were considered particularly relevant to capturing the evolving deep tissue injury process. For patients who remained PI‐free, corresponding variables were extracted from the first 24 h following ICU admission, representing a standardised early clinical baseline prior to prolonged ICU exposure and prior to documented pressure‐related tissue changes. However, because these windows represent different temporal positions within the ICU trajectory, they should not be interpreted as physiologically equivalent observation periods. The framework was therefore used to examine DTPI‐proximal physiological patterns while recognising that direct between‐group comparisons may be influenced by differences in the timing of measurement. Consequently, all haemodynamic comparisons were interpreted within a time‐at‐risk framework rather than as cross‐sectional measurements. Surgical exposure was anchored to the same group‐specific observation windows: the 24 h preceding DTPI identification for patients with DTPI and the first 24 h following ICU admission for PI‐free patients. Hemodynamic instability at ICU admission was treated as a baseline admission characteristic and was not derived from the group‐specific 24‐h time‐at‐risk window.

2.8. Data Domains

The following data elements were collected:

• Demographics and admission characteristics: age, sex, primary ICU admission diagnosis, haemodynamic instability at ICU admission and discharge outcome.

• PI‐related characteristics: DTPI onset timing, anatomical site, device‐related nature and lowest Braden Scale score recorded prior to DTPI development.

• Clinical severity and risk profiles: APACHE II, NRS‐2002 and comorbidities (cardiovascular disease, diabetes, organ failure (specifically heart failure and kidney failure), other chronic diseases such as malignancy, chronic obstructive pulmonary disease, chronic neurological disorders).

• Procedures and critical care interventions: surgical procedure performed within the predefined 24‐h observation window (yes/no), procedure duration standardised to hours, renal replacement therapy, including intermittent haemodialysis and continuous veno‐venous hemofiltration, respiratory support (including mechanical ventilation and oxygen therapy) and extracorporeal membrane oxygenation (ECMO). Specific surgical procedure types were not collected.

• Hemodynamic and laboratory parameters: lowest systolic blood pressure, diastolic blood pressure and mean arterial pressure; presence of sustained hypotension (defined as MAP <65 mmHg) [26]; lowest haemoglobin and albumin levels; presence of oedema; and minimum SpO₂. Oedema (peripheral or generalised) was assessed via clinical inspection and palpation by bedside nurses and WOCNs, coded as a binary variable (present/absent).

• Vasopressor and medication exposure: type (norepinephrine, dopamine, dobutamine, epinephrine and vasopressin), maximum infusion dose (mcg/kg/min) and duration of norepinephrine and dopamine therapy (hours). Because dobutamine, epinephrine and vasopressin exposure was uncommon, these agents were not analysed as separate drug‐specific variables. The highest systolic blood pressure recorded during vasopressor therapy and the simultaneously recorded diastolic blood pressure were also extracted, together with use of sedatives, corticosteroids and anticoagulants. Norepinephrine duration represented the cumulative duration of norepinephrine therapy documented in the medical record and was not restricted to the 24‐h physiological measurement window.

• Antithrombotic exposure: categorised as LMWH, UFH, oral anticoagulants and antiplatelet agents and was operationalised as administration (yes/no) during the group‐specific 24‐h observation window

• Nursing observations: nursing‐derived variables included immobility and oedema. Immobility was defined as explicit bedside nursing documentation that the patient was bedbound and was coded as yes/no independently of the total Braden Scale score. Oedema was coded as present/absent based on clinical inspection and palpation by bedside nurses and WOCNs. Nutritional risk was assessed using the NRS‐2002 total score, with scores ≥ 3 indicating nutritional risk. Overall pressure injury risk was assessed using the total Braden Scale score; individual Braden subscale scores were not analysed separately.

Hemodynamic and laboratory parameters were recorded at least every 4 h. For the time‐at‐risk analyses, the lowest recorded systolic blood pressure, diastolic blood pressure, mean arterial pressure, SpO₂, haemoglobin and albumin values within each predefined 24‐h observation window were used as model inputs rather than mean values.

2.9. Data Preprocessing

Data preprocessing followed the TRIPOD‐AI recommendations. The dataset included both numerical and categorical variables, requiring data type harmonisation prior to model development. Categorical variables such as chronic illness categories, ICU admission diagnoses and anticoagulant types were one‐hot encoded to generate binary indicator features. Duration‐related variables recorded in mixed units (days and hours) were standardised to hours to ensure uniformity. Dialysis exposure was retained as a binary feature (presence vs. absence) to capture clinically meaningful treatment exposure while minimising dimensional complexity. Variables describing PI characteristics were excluded from model development because these data were inherently unavailable for patients without PI and could introduce information leakage; therefore, only pre‐injury time‐at‐risk variables were used for model training. All patients were assigned a binary outcome label (1 = DTPI present, 0 = no DTPI). Missing data (< 5% across all variables) were addressed using median imputation for continuous variables and mode imputation for categorical variables. To prevent data leakage, imputation and standardisation procedures were performed within cross‐validation folds, using only training data. Given the low rate of missingness, continuous variables were standardised using z‐scores prior to model training. Data preprocessing procedures were performed in Python (version 3.10) using the Pandas and Scikit‐learn libraries.

2.10. Feature Selection

Feature selection aimed to identify the most strongly associated variables with DTPI occurrence. To capture both linear and nonlinear relationships, multiple complementary statistical approaches were applied using Python's Scikit‐learn library [27]. Pearson correlation coefficients were used to quantify the strength and direction of linear relationships between continuous variables [28]. Information gain was calculated to evaluate entropy reduction contributed by each feature, identifying variables with the greatest discriminative value for distinguishing DTPI from PI‐absent cases [29]. Chi‐squared testing was applied to categorical predictors and Fisher's score was used to rank features based on between‐class to within‐class variance ratios [30, 31]. To prevent information leakage, feature selection was performed on the training data during cross‐validation. To avoid redundancy, highly correlated features (r > 0.80) were examined and only the variable with greater clinical interpretability was retained. Based on consistency across methods, a subset of top‐ranked predictors was selected for model development (Figures 1, 2, 3; Table 2).

FIGURE 1.

Pearson correlation ranking of features associated with DTPI (N:336). Positive correlations indicate statistical association with DTPI status within the defined time‐at‐risk framework and do not imply causal direction or protective effects. Negative correlations indicate inverse associations. APACHE II = Acute Physiology and Chronic Health Evaluation II; DBP = diastolic blood pressure; LMWH = low‐molecular‐weight heparin; MAP = mean arterial pressure; NRS‐2002 = Nutritional Risk Screening 2002; SBP = systolic blood pressure.

FIGURE 2.

Ranking of DTPI‐related features based on information gain (N:336). Higher information gain values indicate a stronger contribution of each feature to entropy reduction when distinguishing DTPI status within the defined time‐at‐risk framework. Features are presented in descending order of information gain. APACHE II = Acute Physiology and Chronic Health Evaluation II; DBP = diastolic blood pressure; LMWH = low‐molecular‐weight heparin; MAP = mean arterial pressure; NRS‐2002 = Nutritional Risk Screening 2002; SBP = systolic blood pressure.

FIGURE 3.

Fisher's score ranking of features associated with DTPI (N:336). Higher Fisher scores indicate greater between‐class‐to‐within‐class variance ratios for each feature with respect to DTPI status within the defined time‐at‐risk framework and reflect feature relevance rather than predictive performance. APACHE II = Acute Physiology and Chronic Health Evaluation II; DBP = diastolic blood pressure; LMWH = low‐molecular‐weight heparin; MAP = mean arterial pressure; NRS‐2002 = Nutritional Risk Screening 2002; SBP = systolic blood pressure.

TABLE 2.

Performance of machine learning models for DTPI prediction (N:336).

Model	Accuracy (95% CI)	F1 score (95% CI)	Precision (95% CI)	Recall (95% CI)	AUC (95% CI)	Brier score	Calibration slope	Calibration intercept
Logistic Regression	0.921 (0.861–0.970)	0.934 (0.887–0.975)	0.919 (0.857–0.982)	0.950 (0.883–1.000)	0.965 (0.925–0.992)	0.068	0.800	0.126
Decision Tree	0.851 (0.782–0.921)	0.876 (0.814–0.932)	0.869 (0.797–0.944)	0.883 (0.800–0.950)	0.902 (0.834–0.958)	0.085	0.656	0.427
Random Forest	0.921 (0.861–0.970)	0.934 (0.889–0.975)	0.919 (0.859–0.983)	0.950 (0.883–1.000)	0.966 (0.922–0.994)	0.066	1.641	−0.184
XGBoost	0.861 (0.792–0.921)	0.887 (0.828–0.937)	0.859 (0.789–0.933)	0.917 (0.833–0.983)	0.976 (0.946–0.993)	0.076	1.667	−0.587
LightGBM	0.891 (0.822–0.950)	0.911 (0.857–0.959)	0.889 (0.821–0.951)	0.933 (0.867–0.983)	0.969 (0.931–0.994)	0.061	0.708	−0.147
Support Vector Machine	0.921 (0.861–0.970)	0.934 (0.887–0.975)	0.919 (0.857–0.982)	0.950 (0.883–1.000)	0.965 (0.927–0.993)	0.073	1.280	−0.186

Note: Accuracy represents the proportion of correctly classified instances; precision reflects the proportion of true positives among predicted positives; recall represents the proportion of actual positives correctly identified; the F1‐score is the harmonic mean of precision and recall; AUC reflects the model's ability to discriminate between DTPI‐present and PI‐absent patients across classification thresholds.

Abbreviations: AUC = area under the receiver operating characteristic curve; LightGBM = light gradient boosting machine; XGBoost = eXtreme gradient boosting.

The final selected predictors included:

• NRS‐2002 score,

• APACHE II score

• Surgical procedure performed within the predefined observation window (yes/no),

• Dialysis exposure, including CVVH or intermittent haemodialysis (1 = yes, 0 = no),

• Lowest systolic, diastolic and mean arterial pressure (mmHg),

• Chronic disease categories (none, other, organ failure),

• Hemodynamic instability at admission (baseline admission characteristic; MAP <65 mmHg and/or vasopressor requirement),

• Norepinephrine infusion duration (hours) and

• Highest systolic blood pressure recorded during vasopressor therapy (mmHg)

These features were retained for training the ML model, while the remaining variables were excluded. Although 15 predictors constituted the optimal feature set, models were additionally trained using 10‐, 15‐ and 20‐feature configurations to enable comparative performance evaluation.

2.11. Model Development

The dataset was randomly partitioned into training (70%) and testing (30%) sets using stratified sampling to preserve class distribution. Model performance was evaluated using accuracy, precision, recall, F1 score and the area under the receiver operating characteristic curve (AUC‐ROC). Feature importance was derived from ensemble‐based models trained solely on the training data, including Random Forest and eXtreme Gradient Boosting (XGBoost), to identify the most influential clinical predictors of DTPI. Receiver operating characteristic curves and confusion matrices were generated to visualise classification performance. All model development and evaluation procedures followed the TRIPOD‐AI recommendations to ensure transparent and reproducible reporting [23]. All preprocessing, feature selection and hyperparameter optimization were performed exclusively on training data.

2.12. Machine Learning Pipeline

Six supervised ML models were implemented and evaluated: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), XGBoost and LightGBM. Hyperparameter optimization for each model was performed using a 5‐fold cross‐validation grid search on the training dataset. The optimised hyperparameters were as follows: for LR, inverse regularisation strength (C = 0.1) with L2 regularisation; for DT, maximum tree depth of 3 and a minimum of 2 samples required for node splitting; for RF, 200 estimators with feature sampling based on the square root of the total number of features; for SVM, a radial basis function kernel with C = 10; for XGBoost, a learning rate of 0.2, maximum tree depth of 3 and 200 estimators; and for LightGBM, a learning rate of 0.1, 100 estimators and a maximum of 15 leaves per tree. Following hyperparameter optimization, final models were trained on the entire training dataset and evaluated on the independent test set. Model performance was assessed using accuracy, precision, recall, F1 score and the area under the receiver operating characteristic curve (AUC‐ROC).

2.13. Model Evaluation

Model performance was evaluated using accuracy, precision, recall, F1 score and the area under the receiver operating characteristic curve (AUC‐ROC). Receiver operating characteristic curves and confusion matrices were generated to visually assess classification performance across models. Given the moderate class imbalance, evaluation metrics that balance sensitivity and specificity (F1 score and AUC‐ROC) were prioritised. All metric results were computed together with their 95% confidence intervals. In addition, model calibration was examined separately for each model using calibration plots and Brier scores were used to assess agreement between predicted probabilities and observed outcomes. These metrics were used to compare model discrimination, balance between sensitivity and specificity and overall predictive performance.

2.14. Data Analysis

All statistical analyses were performed in Python (version 3.10) using the Scikit‐learn, NumPy and Pandas libraries. Continuous variables were summarised as mean ± standard deviation (SD) for normally distributed data or median with interquartile range (IQR) for non‐normally distributed data. Normality was assessed using visual inspection of histograms and Q–Q plots. Categorical variables were presented as frequencies and percentages. Group differences between DTPI‐present and DTPI‐absent patients were examined using independent samples t‐tests or Mann–Whitney U tests for continuous variables and chi‐square tests or Fisher's exact tests for categorical variables, as appropriate. These analyses were conducted to describe baseline differences between groups and were independent of the ML modelling process; they were not used for feature selection or model training. Statistical significance was defined as p < 0.05 (two‐tailed).

2.15. Ethical Considerations

Ethical approval for this study was obtained from the Koç University Institutional Review Board (IRB No: 2024.339.IRB3.148). The requirement for written informed consent was waived because all data were derived from routine clinical practice and were fully de‐identified prior to analysis, consistent with national regulations and international research ethics standards. No patient‐identifiable information was stored, shared or used during the model development process. To minimise ethical risks in ML applications, several safeguards have been implemented. First, all preprocessing and modelling steps adhered to principles of data minimisation, ensuring that only clinically relevant variables were included. Second, model outputs were used solely for research purposes; no real‐time clinical decisions were influenced by the algorithm and no automated alerts were delivered to clinicians. All procedures complied with the ethical principles outlined in the Declaration of Helsinki and institutional data‐protection standards governing the use of electronic health records in research [32]. Only authorised study personnel had access to the secure research database and all analyses were performed on encrypted, access‐controlled servers.

3. Results

As shown in Table 1, gender distribution differed between groups, with males more common in the DTPI group (61.1% vs. 49.6%, p = 0.04). Patients who developed a DTPI had significantly longer hospital stays (40.9 ± 42.4 vs. 17.6 ± 28.9 days, p < 0.01) and ICU stays (406.7 ± 752.3 vs. 118.7 ± 107.9 h, p < 0.01) than those without PIs. Nutritional risk was higher in the DTPI group (3.9 ± 3.5 vs. 2.4 ± 2.1, p < 0.01), while severity of illness was greater among PI‐absent patients (APACHE II: 30.1 ± 9.8 vs. 26.2 ± 8.8, p < 0.01). Immobility showed a strong association with DTPI (95.7% vs. 76.8%, p < 0.01). Respiratory support use and support‐type distribution did not differ significantly between the groups (97.6% vs. 99.2%, p = 0.418; p = 0.096, respectively). Vasopressor use was universal among PI‐absent patients (100% vs. 51.2%; p < 0.01). Minimum hemodynamic values were consistently higher in the DTPI group—including systolic BP (89.8 ± 15.1 vs. 69.8 ± 17.9 mmHg, p < 0.01), diastolic BP (52.3 ± 10.1 vs. 39.9 ± 12.7 mmHg, p < 0.01), MAP (65.9 ± 11.2 vs. 50.2 ± 13.6 mmHg, p < 0.01) and peripheral oxygen saturation (93.5% ± 5.2 vs. 90.8% ± 7.2, p < 0.01) reflecting the differing time windows used for data extraction. DTPI patients more frequently received LMWH (72.0% vs. 50.4%, p < 0.01) and antiplatelet agents (19.9% vs. 9.6%, p = 0.02). The median time to first document DTPI was 9 days (IQR, 5–18), with the sacrogluteal and foot/ankle regions being the most frequently involved anatomical regions.

TABLE 1.

Clinical and demographic characteristics of patients with and without DTPI.

Variable	DTPI present (n = 211)	PI absent (n = 125)	p value
Demographics
Age, years (mean ± SD)	69.7 ± 14.6	69.1 ± 14.1	0.45
Female sex, n (%)	82 (38.9%)	63 (50.4%)	0.04
Male sex, n (%)	129 (61.1%)	62 (49.6%)
Hospital and ICU stay
Hospital LOS, days	40.9 ± 42.4	17.6 ± 28.9	< 0.01
ICU LOS, hours	406.7 ± 752.3	118.7 ± 107.9	< 0.01
ICU admission category, n (%)
Neurological condition	30 (14.2%)	8 (6.4%)	0.04
Hemodynamic instability	26 (12.3%)	6 (4.8%)	0.04
Infection	10 (4.7%)	16 (12.8%)	0.01
Metabolic/general conditions*	12 (5.7%)	20 (16.0%)	< 0.01
Comorbidities, n (%)
Chronic organ failure	4 (1.9%)	39 (31.2%)	< 0.01
Other chronic diseases	16 (7.6%)	55 (44.0%)	< 0.01
No chronic disease	85 (40.3%)	14 (11.2%)	< 0.01
Scores and clinical measures
Braden score	11.4 ± 2.5	11.2 ± 3.1	0.20
NRS‐2002 (nutrition risk)	3.9 ± 3.5	2.4 ± 2.1	< 0.01
APACHE II	26.2 ± 8.8	30.1 ± 9.8	< 0.01
Immobility, n (%)	202 (95.7%)	96 (76.8%)	< 0.01
Surgical procedure performed	24 (11.4%)	42 (33.6%)	< 0.01
Respiratory support, n (%)
Received	206 (97.6%)	124 (99.2%)	0.418
Not received	5 (2.4%)	1 (0.8%)
Respiratory support type a (n = 206/124), n (%)
Mechanical ventilation	122 (59.2%)	88 (71.0%)	0.096
Conventional oxygen therapy	56 (27.2%)	25 (20.2%)
Advanced/non‐invasive support	28 (13.6%)	11 (8.9%)
Medication exposure
Vasopressor use	108 (51.2%)	125 (100%)	< 0.01
Norepinephrine dose, mcg/kg/min	0.08 ± 0.20	0.27 ± 0.91	< 0.01
Norepinephrine duration, hours	10 ± 11.1	55 ± 43.8	< 0.01
Hemodynamic values
Systolic BP, mmHg	89.8 ± 15.1	69.8 ± 17.9	< 0.01
Diastolic BP, mmHg	52.3 ± 10.1	39.9 ± 12.7	< 0.01
MAP, mmHg	65.9 ± 11.2	50.2 ± 13.6	< 0.01
SpO2 (%)	93.5 ± 5.2	90.8 ± 7.2	< 0.01
Highest systolic BP during vasopressor therapy	130.9 ± 15.2	123.7 ± 18.9	< 0.01
Corresponding diastolic BP at the same measurement	68.7 ± 13.5	69.2 ± 14.6	0.81
Antithrombotic therapy, n (%)
LMWH	152 (72.0%)	63 (50.4%)	< 0.01
UFH	6 (2.8%)	0 (0%)	0.14
Oral anticoagulant	5 (2.4%)	6 (4.8%)	0.37
Antiplatelet agents	42 (19.9%)	12 (9.6%)	0.02
No anticoagulant	45 (21.3%)	52 (41.6%)	< 0.01
DTPI onset and anatomical distribution
Time from ICU admission to first documented DTPI, days, median (IQR)	9 (5–18)	—	—
DTPI identified within the first 14 days, n (%)	144 (68.2%)	—	—
Single anatomical site, n (%)	149 (70.6%)	—	—
Multiple anatomical sites, n (%)	62 (29.4%)	—	—
Most frequently involved anatomical regions, n (%) b
Sacrogluteal region	97 (46.0%)	—	—
Foot/ankle region	57 (27.0%)	—	—
Ear	22 (10.4%)	—	—
Knee/patellar region	16 (7.6%)	—	—
Trochanteric region	13 (6.2%)	—	—

Note: For the DTPI‐present group, physiological values and clinical scale scores represent measurements obtained within the 24 h immediately preceding DTPI onset. For the PI‐absent group, corresponding values reflect the first 24 h following ICU admission. This time‐at‐risk alignment enables comparison across clinically meaningful exposure periods despite the intentionally asymmetric temporal framework. *Metabolic/general conditions include diagnoses such as acidosis, alkalosis and related metabolic disturbances. The category ‘organ failure’ includes heart failure and kidney failure. The category ‘Other chronic diseases’ includes malignancy, chronic obstructive pulmonary disease and chronic neurological disorders. p < 0.05 indicates statistical significance.

Abbreviations: APACHE II = Acute Physiology and Chronic Health Evaluation II; DTPI = Deep Tissue Pressure Injury; ICU = Intensive Care Unit; LMWH = low‐molecular‐weight heparin; MAP = mean arterial pressure; NRS‐2002 = Nutritional Risk Screening 2002; UFH = Unfractionated Heparin.

^a Respiratory support type is presented as a percentage of the total patients who received respiratory support (n = 206 for DTPI group; n = 124 for PI‐absent group). ‘Conventional oxygen therapy’ includes nasal cannula and oxygen mask; ‘Advanced/non‐invasive support’ includes BiPAP, high‐flow nasal oxygen and home ventilator. Surgical procedure performed refers to a procedure occurring within the group‐specific 24‐h observation window and does not indicate the reason for ICU admission. Specific surgical procedure types were not collected. Antithrombotic therapy categories were not mutually exclusive, as patients could receive more than one medication class. Norepinephrine duration represents cumulative treatment duration and was not restricted to the 24‐h physiological measurement window.

^b Patients could have DTPIs at more than one anatomical region; therefore, regional percentages do not sum to 100%.

In Figure 1, LMWH use demonstrated the strongest positive association with DTPI (r = 0.729), followed by norepinephrine duration, which showed a moderate negative association (r = −0.476). Hemodynamic parameters exhibited moderate positive associations, including the lowest MAP (r = 0.466), the lowest diastolic BP (r = 0.402) and the lowest systolic BP (r = 0.392). Surgical procedures (r = −0.256), APACHE II score (r = −0.238) and maximum norepinephrine dose (r = −0.199) showed weak negative associations.

Figure 2 presents the ranking of DTPI features by information gain. LMWH use demonstrated the highest information gain (0.281), followed by norepinephrine duration (0.267) and the lowest MAP (0.131). Surgical procedures (0.035) and steroid use (0.034) demonstrated the lowest information gain values.

Figure 3 presents Fisher's score ranking of features associated with DTPI. LMWH use demonstrated the highest Fisher score (178.676), followed by norepinephrine duration (45.998). Hemodynamic variables, including lowest MAP (43.588), lowest diastolic BP (30.189) and lowest systolic BP (28.458), also exhibited substantial between‐class variance. Lower Fisher scores were observed for the highest systolic BP during vasopressor therapy (7.190) and maximum norepinephrine dose (6.531).

Table 2 presents the predictive performance of six ML models for DTPI classification. Accuracy ranged from 0.851 (DT) to 0.921 (LR, RF and SVM). F1‐scores varied between 0.876 (DT) and 0.934 (LR, RF and SVM). Precision values ranged from 0.859 (XGBoost) to 0.919 (LR, RF and SVM), while recall values ranged from 0.883 (DT) to 0.950 (LR, RF and SVM). AUC values ranged from 0.902 (DT) to 0.976 (XGBoost). Overall, most models demonstrated high predictive performance, with LR, RF and SVM achieving the highest accuracy and F1 scores, whereas XGBoost provided the strongest overall discrimination based on AUC. Calibration performance varied across models. Brier scores ranged from 0.061 for LightGBM to 0.085 for DT, while calibration slopes ranged from 0.656 to 1.667 and calibration intercepts from −0.587 to 0.427 (Table 2; Figure 6).

FIGURE 6.

Calibration curves of machine learning models for DTPI prediction. The dashed diagonal line represents perfect calibration. Curves closer to the diagonal indicate greater agreement between predicted probabilities and observed DTPI frequencies. Brier scores are shown for each model; lower values indicate better probabilistic accuracy.

The ROC curves of the six ML models are shown in Figure 4. All models demonstrated strong discriminative performance, with AUC values ranging from 0.902 (DT) to 0.976 (XGBoost). Substantial overlap among ROC curves indicates generally comparable classification performance across models, although DT showed relatively lower discrimination. Across clinically relevant false‐positive rates, most models achieved high true‐positive rates, supporting their capacity to identify DTPI cases while maintaining relatively low false‐positive rates.

FIGURE 4.

ROC curves of LR, DT, LightGBM, XGBoost, RF and SVM models. The curves display the relationship between the true‐positive rate (sensitivity) and the false‐positive rate (1—specificity). The diagonal dashed line represents random guessing (AUC = 0.5). All models demonstrated strong discrimination, with AUC values ranging from 0.902 to 0.976. DT = decision tree; LGBM = light gradient boosting machine; LR = logistic regression; RF = random forest; SVM = support vector machine; XGB = eXtreme Gradient Boosting.

Figure 5 presents the confusion matrices for the DT, XGBoost (XGB) and Support Vector Machine (SVM) models. For DT, the counts were 33 true negatives, 8 false positives, 7 false negatives and 53 true positives. For XGBoost, the corresponding values were 32 true negatives, 9 false positives, 5 false negatives and 55 true positives. The SVM model demonstrated 36 true negatives, 5 false positives, 3 false negatives and 57 true positives. The horizontal axis represents predicted labels (injured vs. not injured) and the vertical axis represents true labels.

FIGURE 5.

Confusion matrices of the DT, XGBoost and SVM models for DTPI classification. The horizontal axis represents predicted labels (not injured vs. injured), while the vertical axis represents true labels.

4. Discussion

This study identified clinically meaningful differences between ICU patients who developed DTPI and those who remained PI‐free. DTPI occurrence was characterised by greater immobility, higher nutritional vulnerability and prolonged intensive care exposure. Across feature selection approaches, treatment exposures and perfusion‐related indicators consistently emerged as dominant DTPI‐associated variables, particularly LMWH use, norepinephrine infusion duration and minimum blood pressure parameters. Among the evaluated ML models, XGBoost achieved the strongest discriminative performance, with an AUC of 0.976, supporting the potential value of physiology‐informed, time‐window‐based modelling for DTPI risk stratification in intensive care settings.

4.1. Pathophysiological Interpretation of the Findings

The findings demonstrate that patients who developed DTPI differ substantially from PI‐absent patients in both clinical and pathophysiological dimensions, indicating that DTPI is characterised by a distinct early physiological signature. Longer ICU and hospital stay in the DTPI group, accompanied by higher immobility rates and increased nutritional risk, support the interpretation that DTPI represents an injury associated with prolonged exposure to sustained mechanical deformation, perfusion vulnerability and reduced individual tissue tolerance (Table 1). These observations are consistent with previous research identifying sustained mechanical deformation and compromised tissue perfusion as central components of DTPI pathogenesis [1, 14]. The observation that APACHE II scores were lower in patients who developed DTPI may appear paradoxical; however, this finding aligns with the concept of survival bias or the exposure‐time paradox. Patients with relatively less acute physiological instability may survive longer, thereby remaining exposed to prolonged mechanical loading and recurrent perfusion fluctuations that facilitate deep tissue injury [5, 9]. This pattern suggests that DTPI development reflects not only acute illness severity but also cumulative physiological vulnerability over time.

The present findings can be further interpreted within the mechanobiological vicious cycle framework described by Gefen et al. [1]. Sustained mechanical deformation acts as the triggering event, initiating primary, direct deformation damage at the cellular and tissue levels. This damage may subsequently promote secondary inflammatory‐oedema related damage, with increasing interstitial pressure and impairment of microvascular and lymphatic function contributing to tertiary ischaemic damage. Importantly, these processes are cumulative and interactive rather than necessarily linear, because pre‐existing or concurrent ischaemia may amplify direct deformation damage and inflammation‐oedema related damage. Within this framework, immobility may prolong exposure to sustained deformation, whereas haemodynamic instability and impaired perfusion may reduce the capacity of already mechanically stressed tissues to tolerate and recover from such loading. Nutritional vulnerability and underlying comorbidities may further modify individual tissue tolerance and repair capacity. Thus, the predictors identified in this study may represent interconnected components of a mechanobiological injury pathway rather than isolated risk factors [1].

Hemodynamic findings should be interpreted in light of the study's time‐at‐risk framework. Although minimum MAP, systolic and diastolic blood pressure values were relatively higher in the DTPI group than in the PI‐absent group, values in both groups reflected clinically compromised hemodynamic states rather than physiological stability. The pre‐DTPI window was intentionally selected to capture physiological conditions proximal to clinical injury recognition, consistent with contemporary mechanobiological concepts describing deep tissue damage as a time‐dependent process that may evolve before becoming clinically visible [1]. However, because physiological data for PI‐absent patients were obtained during the first 24 h following ICU admission, the two windows represent different stages of the clinical trajectory. Accordingly, the observed hemodynamic differences may partly reflect temporal positioning and severity of illness at the time of measurement rather than DTPI‐specific pathophysiology alone.

Among treatment‐related variables, LMWH use consistently emerged as one of the strongest DTPI‐associated features across Pearson correlation, information gain and Fisher score analyses (Figure 1, 2, 3). Although LMWH has not been implicated as a causal contributor to DTPI development, its prominence likely reflects a clinical profile associated with systemic inflammation, reduced mobility and heightened thromboprophylaxis needs [14, 33]. Thus, LMWH use should be interpreted as an indirect clinical marker of vulnerability rather than a pharmacologic risk factor. Similarly, norepinephrine infusion duration and dose demonstrated substantial discriminatory value. Importantly, this discriminatory contribution should not be interpreted as indicating a positive association with DTPI, because norepinephrine duration and dose showed negative univariate associations. Rather, their prominence in the machine‐learning analyses reflects their contribution to distinguishing between DTPI‐present and PI‐absent patients, irrespective of the direction of association. Within the vicious cycle framework, vasopressor exposure may be clinically relevant as a marker of systemic hemodynamic instability and altered peripheral perfusion rather than as a direct or directionally positive risk factor for DTPI [1, 2, 3, 4, 5]. These findings therefore suggest that vasopressor‐related variables may capture a complex hemodynamic profile relevant to DTPI classification rather than a simple dose‐ or duration‐dependent effect.

4.2. Model Performance and Methodological Implications

All six ML models developed in this study demonstrated strong performance in DTPI classification. Accuracy ranged from 0.851 to 0.921, with Logistic Regression, Random Forest and Support Vector Machine achieving the highest accuracy and F1 scores, while XGBoost demonstrated the strongest discriminative performance with an AUC of 0.976 (95% CI: 0.946–0.993) (Table 2; Figure 4). Calibration performance varied across models, with Brier scores ranging from 0.061 to 0.085 and calibration slopes and intercepts indicating differing degrees of agreement between predicted and observed probabilities (Table 2; Figure 6). These results were higher than the moderate discrimination commonly reported in PI prediction literature (AUC = 0.75–0.80) [16], suggesting that routinely collected nursing and haemodynamic data within clinically defined risk windows may contain clinically meaningful predictive information for early DTPI identification [20, 34, 35]. The overlap of ROC curves and the achievement of high true‐positive rates at clinically relevant false‐positive thresholds suggest that the models maintain strong sensitivity while preserving acceptable false‐alarm levels [17, 36, 37]. Confusion matrix analysis further demonstrated relatively low false‐negative counts, particularly for SVM and XGBoost, supporting their potential utility as early screening tools in critical care populations where missed DTPI detection may have significant clinical consequences [16, 34].

Feature selection analyses based on Pearson correlation, information gain and Fisher score consistently identified a stable set of physiologically meaningful predictors. Treatment‐related indicators (LMWH use and norepinephrine exposure) and perfusion‐related parameters (minimum MAP, systolic and diastolic blood pressure) repeatedly ranked among the most informative features (Figures 1, 2, 3). This pattern aligns with prior conceptualizations of DTPI as a multifactorial injury driven by haemodynamic fragility, mechanical loading and treatment‐related perfusion effects [1, 5, 9, 38]. Notably, model performance remained stable across alternative feature‐set configurations (10‐, 15‐ and 20‐feature sets), indicating that strong predictive accuracy can be achieved using a limited number of routinely documented bedside variables. This observation supports findings from previous methodological studies demonstrating that early clinical data may provide substantial predictive value without reliance on complex EHR aggregation [39, 40, 41]. Moreover, this stability aligns with the principle of feature robustness described in model‐stability literature, suggesting that predictive signals arise from underlying clinical phenomena rather than algorithm‐specific artefacts [42].

Compared with prior ML studies predicting PI risk, which frequently aggregate PI stages and rely on extended laboratory panels or late‐phase physiological data [21, 34, 43], the present study focused exclusively on DTPI and employed a comparator group entirely free of PI. The time‐at‐risk framework, extracting data from the 24 h preceding DTPI onset and the first 24 h following ICU admission for PI‐absent patients, aligns with emerging conceptualizations of DTPI as a time‐dependent physiological process characterised by proximal perfusion deterioration [1, 44]. This design was intended to preserve temporal proximity to DTPI recognition and thereby distinguish DTPI‐specific physiological modelling from conventional admission‐based prediction of general PI risk. At the same time, the asymmetric timing of data extraction limits direct physiological equivalence between groups and should be considered when interpreting the apparent predictive contribution of haemodynamic and treatment‐related variables. Collectively, the findings suggest that temporally structured physiological data may provide useful signals for DTPI risk modelling, although this approach requires prospective evaluation using temporally aligned comparison strategies. These findings should nevertheless be interpreted in the context of the relatively modest sample size and the absence of external validation. Although internal validation demonstrated strong discrimination, model performance may be less stable when applied to different ICU populations or documentation environments. Larger multicentre cohorts are therefore needed to confirm discrimination, calibration and transportability before clinical implementation.

4.3. Clinical Implications for Nursing Practice

A key clinical implication is that DTPI risk signals may be detectable using routinely documented nursing observations of haemodynamics, mobility, nutrition and treatment exposure, when these data are temporally aligned with the relevant risk window. This finding highlights the value of bedside nursing assessments as sources of early physiological information and supports the potential role of nursing documentation in risk stratification processes. Traditional risk assessment tools such as the Braden, Waterlow and Norton scales are primarily designed to identify areas requiring preventive attention rather than to capture rapidly changing PI risk. Integrating temporally structured nursing data may therefore enhance individualised risk recognition in critically ill patients [45, 46, 47]. The time‐at‐risk framework applied in this study is also important when interpreting the haemodynamic findings. Although minimum SBP, DBP and MAP values were relatively higher in DTPI patients than in PI‐absent patients, values in both groups remained within pathophysiologically compromised ranges. This pattern may partly reflect differences in temporal positioning within the clinical trajectory rather than absolute physiological superiority and underscores the importance of interpreting haemodynamic indicators within context‐specific exposure windows.

The prominence of vasopressor exposure, immobility and anticoagulant therapy as influential predictors is consistent with literature describing DTPI as a multifactorial injury initiated by sustained mechanical deformation and driven by cumulative interactions among primary, direct deformation damage, secondary inflammatory‐oedema related damage and tertiary ischaemic damage, with individual tissue tolerance modifying the extent and progression of injury [1, 5, 36, 48, 49]. These findings suggest that nursing surveillance should extend beyond surface skin assessment to incorporate indicators of tissue perfusion fragility and cumulative exposure to sustained mechanical loading. From this perspective, prevention should address not only the magnitude and duration of mechanical loading but also the patient's capacity to tolerate that loading. The results also support individualised preventive strategies. Rather than relying solely on fixed repositioning schedules, nursing interventions may benefit from dynamic risk‐assessment approaches that account for evolving perfusion status, mobility limitations and treatment‐related factors. Such approaches align with emerging recommendations emphasising tissue tolerance‐based repositioning and physiologically informed preventive care [1, 9].

From a nursing informatics perspective, the findings indicate the potential value of decision‐support tools that synthesise routinely collected physiological and treatment data to identify emerging risk patterns [8]. Future integration of similar modelling approaches into digital clinical workflows could facilitate earlier recognition of mechanical and perfusion‐related vulnerability and support timely preventive interventions without replacing clinical judgement [16, 17, 19, 37, 50]. Overall, these findings suggest that DTPI may be conceptualised as a time‐sensitive mechanobiological process characterised by detectable risk patterns prior to visible tissue damage. The study therefore contributes to the development of nursing frameworks centred on individualised tissue tolerance, deep tissue perfusion awareness, temporally informed risk stratification and proactive identification of early physiological signals at the bedside.

5. Limitations

This study has several limitations that should be considered when interpreting the findings. First, the retrospective design intentionally used different time‐at‐risk windows for the comparison groups: physiological data were extracted from the 24 h preceding DTPI onset in the DTPI group and from the first 24 h after ICU admission in the PI‐absent group. Although this approach was selected to capture physiological conditions proximal to DTPI recognition and was informed by the time‐dependent evolution of deep tissue damage, the asymmetric timing may introduce time‐of‐testing bias because the two windows may represent different stages of critical illness. Consequently, observed differences in illness severity, vasopressor exposure, blood pressure levels and stability and ICU length of stay may partly reflect temporal positioning within the clinical trajectory rather than DTPI‐specific pathophysiology alone. The findings should therefore be interpreted as associations within the predefined temporal framework rather than as directly equivalent or causal physiological differences between groups. Second, while haemodynamic and laboratory parameters were recorded at regular intervals, the analysis relied on minimum values within the predefined 24‐h observation windows, which may not fully capture dynamic fluctuations, cumulative perfusion burden or duration of hypotensive episodes. Consequently, transient physiological instability and microcirculatory disturbances relevant to deep tissue injury may have been underestimated. Third, several clinically relevant preventive care variables were inconsistently documented and therefore excluded from analysis, including repositioning practices, use of pressure‐redistribution surfaces, device‐related pressure exposure, staffing characteristics and detailed nutritional intake. The absence of these variables may introduce residual confounding related to care intensity and preventive practices. Fourth, the study was conducted in a single tertiary ICU setting. Although rigorous feature selection and internal validation procedures were applied, the absence of external multicentre validation limits the generalizability and transportability of the models to institutions with different patient populations, staffing models and documentation practices. Fifth, associations identified between treatment‐related variables and DTPI should not be interpreted as causal. Variables such as anticoagulant exposure and vasopressor therapy likely function as markers of clinical complexity, immobility or perfusion vulnerability rather than direct etiologic contributors. Sixth, although the sample size was adequate for exploratory model development and internal comparison of the evaluated algorithms, it remains relatively modest for machine learning model development, particularly when multiple algorithms are trained and compared. This may limit model stability and the precision of performance estimates. External validation in larger, more heterogeneous cohorts is therefore needed to confirm model robustness, calibration and transportability. Seventh, reliance on routinely documented nursing data represents both strength and limitation. Although this approach enhances clinical feasibility, it excludes continuous physiological waveform data that may improve early detection of perfusion compromise. Finally, retrospective performance metrics cannot establish a model's real‐world clinical utility. Prospective implementation studies are needed to evaluate workflow integration, effects on nursing decision‐making, potential alert fatigue, feasibility of real‐time updating and the impact of early warning systems on DTPI prevention outcomes. Although surgical exposure and procedure duration were recorded, specific surgical procedure types were not collected, limiting evaluation of procedure‐specific associations with DTPI.

6. Conclusion

This study demonstrates that DTPI risk may be detectable before visible skin changes using routinely collected nursing and haemodynamic data within a physiologically defined time‐at‐risk framework. Among the evaluated machine learning models, XGBoost provided the strongest discriminative performance, while overall model results consistently indicated that early perfusion‐related indicators, including mean arterial pressure, systolic and diastolic blood pressure patterns, vasopressor exposure, immobility, nutritional vulnerability and selected treatment variables, carry clinically meaningful signals associated with subsequent DTPI development. The findings support conceptualising DTPI as a cumulative and time‐sensitive physiological process rather than a sudden clinical event. Marginal perfusion patterns, prolonged vasopressor exposure, increased immobility and nutritional deficits emerged as a convergent clinical profile preceding DTPI onset. Importantly, the results demonstrate that robust predictive models can be developed with strong internal performance using routinely documented bedside nursing data, without reliance on complex electronic health record aggregation. Clinically, these findings reinforce the limitations of traditional surface‐based risk assessment tools in critically ill populations and highlight the need for physiology‐sensitive approaches to DTPI prevention. Incorporating perfusion‐related indicators into nursing risk assessments may enhance early detection and support individualised preventive strategies. Overall, this study provides evidence that DTPI represents an identifiable and potentially preventable condition when evaluated within an appropriately aligned risk‐time framework. The findings provide an analytical foundation for future development and prospective evaluation of nurse‐centred, explainable decision‐support tools to improve early DTPI prevention in intensive care settings.

Funding

The authors have nothing to report.

Ethics Statement

The study was approved by the Koç University Social Sciences Research Ethics Committee (Approval No: 2024.339.IRB3.148, Approval Date: 03 Oct 2024). Institutional permission to access and use patient records was obtained from the affiliated hospital administration prior to data collection. Due to the retrospective nature of the study and the use of anonymized archival patient data, written informed consent was not required. All data were de‐identified prior to analysis and no personally identifiable information was collected, stored or reported.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.