The hidden dangers in routine medical complaints: uncovering patient harm

Abstract

Background

Patient harm incidents (PHI) significantly impact healthcare quality and outcomes. Despite technological advances, predicting and managing these incidents remains challenging. This study aims to develop a predictive model to distinguish the genuine PHI from medical complaints and non-harmful events.

Methods

A retrospective study was conducted using data collected from January 2014 to December 2023, encompassing patient interactions, treatments, and complaints in the authors’ institution. Variables considered included demographic details, clinical factors, and complaint characteristics. The predictive model was developed using least absolute shrinkage and selection operator (LASSO) regression and validated via a split-sample method. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration plots, the Hosmer-Lemeshow test, and decision curve analysis.

Results

The study included 987 medical complaints, of which 361 involved PHI. Using LASSO and logistic regression analyses, the model identified key predictors of PHI including the choice of treatment methods, healthcare providers’ professional behavior, and the nature of the complaints. The model achieved an AUC of 0.917 (95% CI: 0.895–0.939) in the training set and 0.904 (95% CI: 0.870–0.938) in the test set. Model performance was further supported by calibration and decision curve analysis results.

Conclusion

The predictive model shows promise in identifying PHI from service complaints within our hospital. Key predictors, such as treatment decisions and healthcare providers’ professional conduct, appear to play a notable role in patient safety. By utilizing this model, healthcare facilities may enhance their ability to identify and address factors that could contribute to PHI.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12913-025-13105-w.

Introduction

Globally, medical safety is a critical issue within the public health sector. The World Health Organization estimates that millions of patients worldwide suffer from preventable harms due to adverse events in medical processes each year [1]. These incidents contribute significantly to the global disease burden, with impacts comparable to severe diseases such as tuberculosis and malaria [2].

As medical services increase in complexity and patient demands diversify, hospitals face challenges in effectively differentiating and managing medical complaints to identify patient harm incidents (PHI) [3]. The medical complaint is defined as any report of dissatisfaction with the service or care provided, which can be initiated by patients or their relatives [4]. PHI is defined as any instance where a patient suffers from temporary or permanent impairment, physical or psychological injury, or death, as a result of their medical care rather than from the underlying condition for which they are being treated [5]. Research indicates that a robust complaint management system can boost patient satisfaction and substantially reduce legal risks and financial losses for hospitals [6–9]. Despite their potential value, medical complaints are often managed procedurally rather than analytically. Previous research has largely focused on complaint frequency and typology, with limited attention to how complaint narratives might reveal PHI [10]. In real-world hospital settings, PHI is frequently embedded within complaints that appear routine on the surface. Patients or their families may not recognize that harm has occurred, or may choose to express it through general dissatisfaction [11]. Therefore, identifying PHI from routine complaints requires structured analysis and clinical interpretation.

From a patient safety perspective, distinguishing PHI-related complaints from non-harm complaints enables early identification of high-risk patterns. Theoretical frameworks such as Reason’s Swiss Cheese Model and Donabedian’s Structure-Process-Outcome framework emphasize that harm often results from accumulated system-level vulnerabilities rather than isolated mistakes [12, 13]. Applying these frameworks to complaint data can help uncover subtle yet recurring system vulnerabilities. Consequently, systematically distinguishing PHI from non-PHI cases within routine hospital complaints provides a critical opportunity to better understand the underlying mechanisms of medical harm and to inform more effective institutional safety responses.

This study aims to develop a predictive model by analyzing hospital complaint data to distinguish complaints involving actual PHI from those merely expressing dissatisfaction with services. The model incorporates perspectives from both physicians and patients, analyzing factors such as physicians’ professional behavior, patients’ baseline conditions, and the complexities of the treatment process that can influence the occurrence of PHI. The study utilizes internal hospital data and nomogram charts to better identify factors contributing to PHI. This approach aims to provide evidence-based support for hospital management decisions, potentially improving healthcare quality and patient safety. The approach is designed from the dual perspectives of doctors and patients, recognizing that PHI are a shared concern requiring mutual engagement and understanding. By using nomogram charts for data visualization, the study provides a clearer understanding of the factors associated with PHI.

Methods

Study design and data source

This retrospective study examined medical complaints recorded over a decade, from January 2014 to December 2023. The hospital, a pioneering institution nationally recognized for its certification, supports a large capacity with thousands of beds. Annually, it handled several million outpatient and emergency visits, hundreds of thousands of inpatient discharges, and over a hundred thousand surgical operations. To accurately categorize medical complaints as harm events, our approach involved a two-tier process: Initially, all complaints were reviewed by a team of trained medical professionals, including clinicians and patient safety officers, who applied predefined criteria to identify potential harm events based on the complaint descriptions [14, 15]. A complaint was flagged as indicating possible PHI if it met one or more of the following criteria:

1. explicit mention of physical or psychological injury following medical care;

2. allegations of diagnostic error (e.g., misdiagnosis, delayed diagnosis);

3. treatment-related incidents (e.g., medication errors, surgical complications, procedural mishaps);

4. references to unplanned readmissions, emergency interventions, or intensive care transfers;

5. any death or permanent disability that could be reasonably attributed to medical care rather than the underlying disease.

Subsequently, those identified in the initial screening underwent a rigorous validation process involving a detailed review of the patient’s medical records, interviews with the involved healthcare providers, and consultations with external medical experts as necessary. The data encompass detailed records of patient and physician interactions, treatments, and subsequent complaints filed during this period. To address missing data, multiple imputation methods were employed, enhancing the robustness of the analyses. Part of the data collection included peer evaluations of the involved physicians, where randomly selected doctors from the same department provided anonymous assessments, ensuring an unbiased sample. All personal identifiers were replaced with unique codes to maintain anonymity, as per the ethical guidelines of our institution. All procedures and methodologies were rigorously designed to adhere to ethical standards, and the study received approval from the appropriate institutional review boards (2024PS1221K).

Variable selection and measurement

All candidate variables were selected based on clinical relevance, prior literature, and data availability within the complaint records [16, 17]. Each variable was operationally defined using standardized coding procedures to ensure consistency across evaluators and time points. Variables were excluded if they exhibited excessive missingness, strong collinearity, or ambiguous definitions, to preserve model robustness and interpretability.

Demographic and socioeconomic factors

Variables include patient age, recorded in years to explore age-related trends in safety incidents. Gender, categorized as male or female, and insurance type, classified into insured or self-paying, are analyzed to assess their effects on healthcare access and outcomes. In the study country, not all individuals or diseases are fully covered by insurance; social insurance generally only extends to the most basic diseases and treatments. Consequently, coverage for certain diseases or patients might be fully funded by insurance, while others may require partial or full out-of-pocket payments by the individuals themselves. Consultation Fee is included to evaluate its impact on the utilization of healthcare services.

Clinical and health status factors

Clinical variables encompass the timing of treatment, categorized by year, month, and season quarter to identify timing effects on patient safety. The number of hospitalizations reflects healthcare utilization. Examination complexity is assessed using a three-tier scale (easy, medium, and hard). This scale is defined uniformly across all medical specialties by criteria such as the estimated duration of the examination, the technical skills required, and the potential for complications (Table 1).

Table 1.

Criteria for the Three-Tier examination complexity scale

Complexity Level	Typical Chief Complaints	Required Workup & Management	Consultation Duration & Resource Use	Potential for Urgent Intervention
Easy	Simple, clearly defined problems (e.g., sore throat, minor injury, medication refill)	Physical exam only or basic point-of-care tests (e.g., blood pressure, rapid strep test); no treatment or simple prescriptions	Brief (≤ 10 min); no advanced diagnostics; no specialist involvement; minimal resource use	Very unlikely
Medium	Vague, moderate, or multiple complaints (e.g., abdominal pain, dizziness, persistent cough)	Routine labs or imaging (e.g., blood tests, urinalysis, chest X-ray); symptom-specific treatment	Moderate (10–20 min); may require basic imaging, repeat vitals, or short-term follow-up	Possible but manageable on-site
Hard	Severe, nonspecific, or potentially life-threatening symptoms (e.g., chest pain, confusion, trauma)	Advanced diagnostics (e.g., ECG, CT scan, lumbar puncture); urgent therapeutic decisions	Long (≥ 20 min); high resource use; likely specialist referral or hospital admission	High risk; requires immediate action review

CBC complete blood count, ECG electrocardiogram, CT computed tomography, MRI magnetic resonance imaging, POC point-of-care

Service accessibility is rated as either ‘convenient’ or ‘inconvenient’, based on factors like waiting times, location accessibility, and appointment availability (Table 2). The type of visit is differentiated as ‘emergency’ or ‘routine’, each associated with distinct patient experiences and outcomes.

Table 2.

Criteria for the Two-Tier service accessibility classification

Category	Definition	Key Factors Considered	Typical Visit Type & Patient Experience
Convenient	Healthcare services that are timely, accessible, and predictable	- Waiting time < 15 min - Facility within 5 km - Appointment available within 24–48 h	Mostly associated with routine visits; patients experience minimal delays, consistent care, and satisfaction
Inconvenient	Services that are delayed, difficult to reach, or poorly scheduled	- Waiting time > 30 min - Facility > 10 km away - Appointment delays > 3 days or unavailable	More frequent in emergency visits or rural settings; patients often report uncertainty and logistical burden

Classification was based on patient-reported experiences and administrative records

Costs of visits and the complexity of treatment plans are also classified using the easy, medium, and hard scale. The criteria for this classification are standardized for each specialty by a committee of interdisciplinary experts who define the complexity based on the number of procedural steps, the necessity for specialized equipment, and the level of interdisciplinary coordination required.

Treatment methods are categorized as surgical or non-surgical. Lifestyle habits are noted as smoking or non-smoking. Comorbidity presence is determined by a binary yes or no, depending on the presence of one or more significant chronic conditions that could affect treatment outcomes. Patient compliance is also recorded as yes or no, based on adherence to treatment plans as documented in follow-up visits.

Disease type is categorized as acute or chronic, and severity is classified as mild, moderate, or severe, with clear definitions provided by clinical guidelines to ensure consistent assessment across specialties. Intensive Care Unit (ICU) admissions, the frequency of the condition, stays exceeding 30 days, and discharge outcomes (Cured, Improved, Dead) are accurately tracked. The gender match between patient and doctor is also recorded, investigating potential biases in patient-provider interactions.

Professional and operational factors

Variables related to healthcare providers include physician gender, professional title, department, years of experience, and workload. Skills and behavioral attributes such as communication skills, cooperation level, team skills, professional ethics, emotional stability, acceptance of new technology, leadership, decision-making style, and relationship with colleagues are evaluated using a structured assessment tool. This tool comprises a standardized questionnaire rated on a Four-point Likert scale, ranging from ‘Very good’ to ‘Weak’. For detailed scoring and the associated tables, see Supplementary text. Pilot testing was conducted on a sample of 50 healthcare providers. Internal consistency reliability was assessed using Cronbach’s alpha, which yielded a coefficient of 0.71, indicating acceptable internal consistency. Exploratory factor analysis supported construct validity, revealing a unidimensional structure accounting for over 50% of the total variance.

Each healthcare provider was assessed across nine behavioral attributes using three sources: one self-evaluation, one peer evaluation (from a randomly assigned colleague within the same department), and one qualitative feedback entry from either a nurse or a patient. Final scores were computed as a weighted average of these three components: peer evaluation (50%), self-evaluation (30%), and narrative-based supervisor-coded nurse/patient feedback (20%). To ensure objective quantification of the qualitative narratives, trained supervisors used a coding rubric structured as a Behaviorally Anchored Rating Scale (BARS). This method is designed to enhance rating accuracy by anchoring each point on the scale with specific, observable behavioral descriptors [18]. For each attribute, the BARS defined performance at four levels: 4 (consistently exceeds expectations), 3 (meets expectations), 2 (requires improvement), and 1 (significant concerns). This structured protocol minimized rater bias and ensured consistent conversion of qualitative feedback into a reliable score. The final composite scores were rounded to the nearest integer (1–4).

A total of 987 complaint incidents were collected, involving 680 unique healthcare providers. Given that some providers were associated with multiple complaints, the number of evaluations was based on unique providers rather than incidents. Self-evaluations achieved 100% completion (680/680), while 850 peer evaluation requests were needed to reach 680 responses (80% yield). This approach highlights the department’s commitment to maintaining high standards of care and accountability, ensuring that all evaluations are thorough and reflective of each provider’s performance over the evaluation period.

Complaint characteristics

Complaint data are categorized by location (outpatient, emergency, ward), whether they occurred in the main hospital area or a branch campus, and the mode of complaint (in-person or by telephone). The involvement of the surgery department (yes or no) and “complaint subject” (doctor, nurse, other staff) are also recorded. Complaints are classified into categories such as medical care, nursing care, medication, service, process, and others, and are analyzed based on the timing of their occurrence, whether during a working day or weekend.

Model development and validation

For the development and evaluation of the nomogram, the dataset was randomly partitioned into a training set and a validation set, comprising 70% and 30% of the participants, respectively. The training set was utilized to identify the most relevant predictors using the least absolute shrinkage and selection operator (LASSO) regression technique [19]. The significant predictors that emerged were further examined using multivariate logistic regression, considering a significance level of P < 0.05 for inclusion in the model. LASSO regression was used to reduce multicollinearity and select the most informative predictors from a large set of candidate variables [20]. This method is particularly suitable for high-dimensional data and helps prevent overfitting. Significant predictors identified by LASSO were subsequently included in a multivariable logistic regression model, with a threshold of P < 0.05 for retention. Logistic regression was selected for its interpretability and compatibility with binary outcomes. A nomogram was then developed to provide a user-friendly tool for individualized risk estimation, consistent with the study’s applied objective.

Model performance evaluation

The performance of the nomogram was evaluated by its receiver operating characteristic (ROC) curve, with the area under the curve (AUC) serving as the measure of predictive accuracy. The model’s calibration was assessed using the Hosmer-Lemeshow test and represented visually via calibration plots [21]. Additionally, the clinical utility of the nomogram was explored through clinical decision curve analysis (DCA) and clinical impact curve (CIC), with the optimal threshold for clinical application being identified by the highest Youden index [22]. These validation processes, including assessments of discriminative ability and model calibration, were replicated on the validation set. The optimal model was encapsulated as an open-source and free predictive web-based application through the “shiny R” program package.

Statistical analysis

Continuous variables were described as either mean ± standard deviation (SD) or median with interquartile range (IQR), depending on their distribution. Group comparisons for these variables were performed using the t-test or the Mann-Whitney U test, depending on the distribution characteristics of the data. Categorical variables were summarized as frequencies (percentages) and group differences were evaluated using chi-square tests. Use 10-fold cross-validation to verify the optimal tuning parameter (λ) for LASSO regression analysis [23]. We further examined variables that were excluded from the final model due to poor data quality, lack of statistical significance, or collinearity. Multicollinearity among predictors was assessed using the variance inflation factor (VIF). The final logistic model included only predictors that were retained by LASSO and demonstrated statistical significance (P < 0.05), ensuring a parsimonious yet clinically meaningful nomogram. All statistical analyses were conducted using R, version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Characteristics of participants

In this study, we analyzed a total of 987 cases, categorizing them into two groups based on the occurrence of PHI to assess their baseline characteristics. The group experiencing PHI included 361 cases, whereas the group with No-PHI comprised 626 cases. The comparative analysis revealed no significant difference in gender distribution, with males representing 43.5% and females 56.5% in the PHI group, and 41.5% males and 58.5% females in the No-PHI group (P = 0.594). Notable differences were observed in insurance coverage, with a significantly higher percentage of cases with health insurance in the PHI group (93.6%) compared to the No-PHI group (80.8%; P < 0.001).

In terms of examination complexity, 30.2% of the PHI group faced hard examinations, significantly higher than the 21.4% in the No-PHI group (P < 0.001). The complexity of the treatment plan also differed significantly between the groups, with the PHI group having a higher proportion of complex treatments (26.9% vs. 11.0% in the No-PHI group; P < 0.001). Lifestyle factors such as smoking were more prevalent in the PHI group (20.5% vs. 13.1%; P = 0.003), and comorbidity complexity was notably higher (22.7% vs. 14.1%; P = 0.002).

Furthermore, the PHI group had significantly higher instances of acute disease severity (24.4% vs. 9.6%), more frequent admissions to the ICU (14.7% vs. 2.4%), and longer hospital stays over 30 days (10.8% vs. 2.1%), all with P < 0.001. For detailed data, see Supplementary Table 1.

Association of candidate predictive variables

An initial analysis was conducted to determine the AUC values for each potential predictor, subsequently ranking these values in descending order. The highest AUC value was observed for Communication Skills of doctors at 0.8279, indicating a strong predictive capability. This was closely followed by Cooperation Level and Leadership, with AUC values of 0.8157 and 0.7692, respectively. Other predictors such as Decision-Making Style, Teamwork Skills, and Acceptance of New Technologies also showed considerable discriminative power, each with AUC values above 0.75 (Fig. 1).

Fig. 1.

ROC and AUC Values for Predictors. A This image represents the single-factor ROC curve (B)This figure presents the Area Under Curve (AUC) for several predictors, illustrating their diagnostic accuracy in a predictive model

Significantly, the assessed variables of patients, including Consultation Fee and Treatment Method Choice, exhibited AUC values exceeding 0.5, suggesting their relevance in the predictive model. However, variables such as Patient Gender, Compliance with Doctor’s Orders, Number of Hospitalizations, and Lifestyle Habits recorded AUC values below 0.5, indicating weaker predictive strength.

The LASSO regression method pinpointed 18 variables with significant coefficients, specifically Year of Visit, Number of Times Condition Explained, Treatment Plan Complexity, Treatment Method Choice, Disease Severity, Admission to ICU, Hospital Stay Over 30 Days, Professional Title, Communication Skills, Cooperation Level, Professional Ethics, Decision Making Style, Relationship with Colleagues, Place of Complaint, Mode of Complaint, Surgery Department Type, Nature of Complaint, and Main Hospital Area. These variables demonstrate a significant predictive power in the model (Fig. 2). Given the overall acceptable range of correlations among variables (Fig. 3), a multifactor analysis was initiated to identify the key predictors of outcomes.

Fig. 2.

LASSO Regression for Feature Selection. This figure depicts the process of selecting important features using LASSO regression in two sections: (A) displays the paths of the coefficients as they change with different logarithmic values of lambda, pinpointing non-zero coefficients at the best lambda as essential features; (B) demonstrates the determination of the best lambda through 10-fold cross-validation, indicated by the criterion’s minimum value plus one standard deviation. This technique efficiently identifies the most significant demographic and clinical predictors for the model

Fig. 3.

Heatmap of Variable Correlations. This figure presents a heatmap utilizing bubble sizes and colors to depict the strength and direction of correlations between variables

Characteristics of participants in the train and test groups

According to a 7:3 random allocation rule, 690 participants (70%) and 297 participants (30%) were randomly assigned to the training and test sets, respectively. A detailed comparison between the training and test sets is provided in Supplementary Table 2. No significant differences were detected between the two groups (P > 0.05), suggesting that both sets are well-balanced and provide good validation efficacy.

Prediction of PHI in the train and test groups

In a multifactorial logistic regression analysis, the odds ratios (OR) were determined for specific variables relative to their reference categories. Non-surgical treatment methods, compared to surgical approaches, had an OR of 0.195, indicating a significantly lower likelihood (95% Confidence Interval [CI]: 0.128–0.297, P < 0.001). Weak communication skills, as opposed to very good ones, increased the odds (OR: 4.606, 95% CI: 3.047–6.959, P < 0.001). Similarly, weak cooperation levels and decision-making styles were also associated with higher odds of PHI compared to very good levels, with OR of 3.709 (95% CI: 2.331–5.903, P = 0.001) and 3.046 (95% CI: 2.028–4.570, P = 0.001), respectively. Conversely, a good relationship with colleagues significantly reduced the odds (OR: 0.130, 95% CI: 0.078–0.218, P < 0.001) compared to weaker relationships. Complaint modes showed that telephone complaints were less likely than in-person visits, with an OR of 0.259 (95% CI: 0.165–0.407, P = 0.001). Non-medical nature of complaints, compared to medical ones, reduced the likelihood (OR: 0.455, 95% CI: 0.328–0.630, P = 0.001). Lastly, being in a non-main hospital area, as opposed to the main hospital area, showed a decreased risk (OR: 0.387, 95% CI: 0.254–0.590, P = 0.007). All variables in the final logistic regression model had VIF values < 5 (range: 1.01–4.34), indicating no significant multicollinearity. These significant factors were integrated to construct the predictive nomogram, enhancing the model’s utility in predicting outcomes based on these adjusted comparisons (Fig. 4).

Fig. 4.

Nomograms for Predicting Patient Safety Incident. This figure showcases the nomograms used to estimate patient safety in the development group. Significance levels are indicated as **P < 0.01, *P < 0.001

The efficacy of the nomogram was evaluated using ROC curve analysis across several cohorts. In the training group, the AUC was 0.917 (95% CI: 0.895–0.939), illustrated in Fig. 5A. For the test group, the AUC improved to 0.904 (95% CI: 0.870–0.938), as shown in Fig. 5B. These outcomes highlight the nomogram model’s excellent capacity for discrimination and its predictive accuracy. At the maximum Youden index, the optimal cut-off value in the training set was 0.74. Calibration curves for both cohorts showed that the model’s predicted probabilities of PHI align closely with actual outcomes, forming an approximate 45-degree angle (Fig. 6A and B). This alignment was further supported by the Hosmer-Lemeshow test, which indicated no significant lack of fit, with p-values of 0.615 in the training cohort. The DCA and CIC results indicate a strong agreement between the model’s predictions and observed realities (Fig. 6C and D).

Fig. 5.

ROC Curves for Patient Safety Incident Prediction. This figure demonstrates the performance of patient safety incident prediction models. Part (A) displays the ROC curve derived from the training dataset, while part (B) shows the ROC curve for the external validation dataset

Fig. 6.

Nomogram Calibration Curves and Clinical Decision and Impact Curve Analysis This figure displays calibration curves for the nomogram, where (A) represents the training set and (B) depicts the validation set. These curves assess the accuracy of the predicted probabilities in comparison to the actual outcomes. C shows the Decision Curve Analysis (DCA), where the x-axis represents the range of threshold probabilities and the y-axis illustrates the net benefit calculated relative to the default strategies of treating all or none. D displays the Clinical Impact Curve (CIC), which details the number of patients who would be flagged by the model across various threshold probabilities (x-axis) and shows how many of these flagged cases are true positive events (y-axis)

Web-based application

We used the “R Shiny” package to create a visual and operational interface for the model. Users can directly obtain predicted PHI probabilities by entering or selecting variables in the web-based application (https://shengjinghospital.shinyapps.io/Model/).

For detailed instructions on usage, refer to Supplementary Fig. 1.

Discussion

To the best of our knowledge, this study is among the first to apply a nomogram to categorize and manage medical complaints. This model aims to organize medical complaint data and predict incidents associated with patient harm. While the results show potential for providing useful insights to hospital decision-makers and department leaders regarding the occurrence of PHI, we acknowledge that the findings are preliminary and should be interpreted with caution. The model can contribute to understanding factors that may influence the occurrence of PHI, but further validation is necessary.

The baseline characteristics revealed significant differences between the PHI and No-PHI groups. Patients in the PHI group were more likely to be insured, have complex examinations and treatment plans, and suffer from severe diseases requiring ICU admission or extended hospital stays [24, 25]. This finding aligns with Panagioti et al. [26], who emphasized that older patients with chronic conditions face higher risk of safety incidents. Furthermore, insured patients tended to receive more complex examinations and treatments, possibly reflecting more aggressive medical management that carries higher procedural risks—a pattern similarly reported by Scott et al. [27]. ICU admissions and prolonged hospital stays in the PHI group suggest higher baseline severity and complications, consistent with Pagnamenta et al.’s findings that more intensive care environments are associated with higher adverse event rates [28]. For medical providers, PHI cases were more frequent in surgical departments and involved male doctors with advanced titles, possibly due to the complexity of cases they handle [29]. Notably, soft skills such as communication, cooperation, and teamwork were rated lower among providers in the PHI group [30]. These findings underscore that both patient-related factors (like disease severity and lifestyle habits) and provider-related factors (such as professional experience and interpersonal skills) contribute to the likelihood of PHI occurrence [31, 32]. Alsabri et al. [33]and Wekker et al. [34]have emphasized similar findings, suggesting that poor team dynamics contribute to higher error rates. Our study reinforces these observations in a real-world complaint dataset, highlighting the value of peer-assessed soft skills as predictive signals for patient safety incidents—an area not commonly captured in structured electronic health record data.

The significance of our findings lies in highlighting the critical role of environmental and human factors in patient safety outcomes. High-traffic hospital areas with complex operations were identified as higher-risk zones for safety incidents, aligning with studies by Zhang et al., which emphasize how hospital architecture and high-stress environments contribute to PHI [35, 36]. Additionally, Bernhardt’s review emphasizes the significance of the healthcare environment in supporting overall healthcare safety, further underlining the importance of architectural considerations in main hospital areas [37]. These insights point to the necessity of optimizing hospital design and operations to enhance patient safety [38, 39].

The nature and mode of patient complaints emerged as significant predictors in our model, indicating that how and what patients choose to complain about can provide valuable insights into potential safety issues [40]. This finding aligns with the work of Reader et al., who demonstrated correlations between complaint types and adverse patient outcomes similar to our PHI definition [6, 40]. Utilizing patient feedback as an early warning system can thus aid healthcare institutions in proactively addressing safety concerns [41].

Interpersonal factors emerged as pivotal in preventing medical errors and enhancing patient care quality [42]. Effective communication and strong relationships among healthcare staff, as highlighted in the research by Leonard et al. and Beuzekom et al., are essential components of a safe healthcare environment [43, 44]. Our study reinforces this by showing that positive team dynamics significantly impact patient outcomes. This underscores the importance of ongoing training and organizational support to foster collaborative and efficient work environments that prioritize patient safety.

Additionally, the appropriateness of treatment method choices plays a significant role in patient safety, echoing findings from Greenberg et al. and Bright et al., who demonstrated a strong correlation between clinical decisions and PHI [45, 46]. This emphasizes the need for evidence-based medical decision-making and continuous professional development to reduce the likelihood and severity of adverse events [47].

The predictive model demonstrated strong performance within our dataset, achieving a training accuracy of 0.917 and a testing accuracy of 0.904, with tight confidence intervals (95% CI: 0.895–0.939 for training and 0.870–0.938 for testing). This suggests the model’s potential effectiveness in classifying patient complaints in our hospital setting. By accurately identifying and categorizing complaints based on their potential impact, the model can assist in refining complaint-handling protocols and addressing systemic issues that contribute to recurring complaints [48]. This may help prevent potential harm and optimize the use of hospital resources, ensuring timely and effective interventions. Although the model is labeled as “predictive,” its application is inherently reactive. Several of the variables—such as complaint content and mode—are only available after the event has occurred, making the model unsuitable for prospective harm forecasting. Nevertheless, in practice, it offers significant value for post-hoc risk detection: it can help hospitals prioritize high-risk complaints, uncover hidden safety threats embedded in routine service feedback, and support targeted quality improvement initiatives. From a research perspective, future studies may explore incorporating real-time clinical variables or patient-reported outcomes to enhance the model’s ability to function as a proactive safety surveillance tool.

All in all, this model highlights the importance of interpersonal dynamics and effective communication in healthcare. Key variables include ‘Nature of Complaint’ and ‘Mode of Complaint’, which are crucial for identifying and addressing severe medical issues and ensuring complaints are handled promptly. Enhancing ‘Relationship with Colleagues’ and ‘Communication Skills’, particularly among doctors, is vital. Managers should focus on communication training that fosters understanding, trust, and patient safety. This includes effective listening, clear explanations of procedures, and empathy training, which are essential for improving patient outcomes and reducing complaints and PHI [49].

This study has some limitations. Firstly, the data originates from a single center, which could limit the generalizability of the findings and underscores the need for external validation to assess the applicability of the results across different settings. Secondly, issues related to data privacy and ongoing legal cases may restrict the availability and sharing of comprehensive incident data, potentially leading to selection bias and partially skewing the analysis. Thirdly, the retrospective nature of the data collection might introduce inaccuracies and inconsistencies in how variables are defined and recorded, possibly affecting the reliability and reproducibility of the study’s conclusions. In the future, it is hoped that more comprehensive datasets can be included to enhance the robustness and scope of the analysis. Additionally, this study predominantly utilizes categorical variables, which may limit the depth and explanatory power of the findings, particularly in analyzing disease severity and cost implications. Lastly, this study categorizes complaints into specific areas such as medical care, nursing care, and medication, among others. However, the complex nature of healthcare often results in overlapping categories of complaints.

Conclusion

This study developed and validated a nomogram-based model to differentiate PHI from routine medical complaints. Among 987 complaints, our analysis identified key predictors of PHI, including treatment method, ICU admission, disease severity, physician communication and cooperation skills, and the nature and mode of the complaint. The model achieved high predictive accuracy, and these findings suggest that specific interpersonal, procedural, and contextual factors are critical in anticipating patient harm. Beyond prediction, the study highlights the value of real-world complaint data in identifying hidden safety risks and underscores the significance of soft skills and environmental factors in hospital safety culture. The developed web-based tool may aid healthcare administrators in early identification of high-risk complaints and support targeted quality improvement efforts.

Abbreviations

PHI

Patient Harm Incident

No PHI

Non–Patient Harm Incident

ICU

Intensive Care Unit

SD

Standard Deviation

AUC

Area Under the Curve

LASSO

Least Absolute Shrinkage and Selection Operator

ROC

Receiver Operating Characteristic

CI

Confidence Interval

SPSS

Statistical Package for the Social Sciences

OR

Odds Ratio

Authors’ contributions

Conceptualization, Xueting Chen and Xinyu Wen; Methodology, Shaoting Luo; Validation, Yong Zhang; Formal Analysis, Shaoting Luo; Data Curation, Xinyu Wen, Boyu Yao, Cui Wang, Qingbing Li, and Wei Wang; Writing– Original Draft Preparation, Shaoting Luo and Xueting Chen; Writing– Review & Editing, Shaoting Luo, Xueting Chen, Yong Zhang, and Lianyong Li; Supervision and Funding, Yong Zhang and Lianyong Li.

Funding

This work was supported by the “2023 Applied Basic Research Program of Liaoning Province” (2023JH2/101300022) and the “Planting Plan” Project for Clinical Research of Shengjing Hospital.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the principles of the Declaration of Helsinki. This study was approved by the Ethics Committee of Shengjing Hospital of China Medical University (Approval Code: 2024PS1221K). Given that this was a retrospective study and all data were anonymized, the Ethics Committee determined that there was no risk of privacy breaches, and as a result, informed consent was waived.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.