Determinants of patient satisfaction in a primary healthcare facility in Bangladesh

Abstract

Background

Patient satisfaction is an important indicator of healthcare quality and system responsiveness, particularly in primary healthcare systems of low- and middle-income countries. In Bangladesh, empirical evidence on determinants of patient satisfaction in public primary healthcare facilities remains limited. This study examined factors associated with patient satisfaction in a public primary healthcare facility in Bangladesh and explored predictive patterns using machine-learning models to complement conventional analysis.

Methods

A cross-sectional study was conducted among 508 patients attending the Savar Upazila Health Complex between April and May 2025. Although conducted in a single upazila-level public healthcare facility, the setting reflects typical primary healthcare service delivery in Bangladesh. Patient satisfaction was assessed using the Patient Satisfaction Questionnaire Short Form (PSQ-18). Sociodemographic, patient-related, and facility-related variables were collected through structured interviews. Bivariate analyses and multivariable binary logistic regression were performed to identify determinants of higher patient satisfaction. In addition, six machine-learning models were applied to evaluate predictive performance and identify influential predictors. Models were tuned using five-fold stratified cross-validation optimizing ROC-AUC and evaluated on a held-out test set.

Results

The mean patient satisfaction score was 64.23 (SD = 8.86; score range: 18–90). Based on Bloom’s cutoff criteria, 74.2% of respondents were classified as having moderate–good satisfaction (≥ 60% of the possible score range). Multivariable logistic regression identified several factors significantly associated with higher satisfaction, including transportation convenience (AOR = 2.54, 95% CI: 1.04–6.19), comfortable waiting-room seating (AOR = 2.18, 95% CI: 1.12–4.24), availability of healthy food (AOR = 2.48, 95% CI: 1.25–4.93), receipt of free medicines (AOR = 2.35, 95% CI: 1.18–4.69), dental-related conditions (AOR = 5.23, 95% CI: 1.13–24.26), and nursing staff behavior (overall association, p < 0.001). Among the machine-learning models, Random Forest demonstrated the strongest discrimination (ROC-AUC ≈ 0.82), while XGBoost achieved the highest classification balance (F1-score ≈ 0.87). Feature-importance analysis highlighted interpersonal care and facility-related factors as dominant predictors of patient satisfaction.

Conclusion

Patient satisfaction in primary healthcare settings is influenced by accessibility, facility environment, and interpersonal care. Within the studies upazila-level context and similar primary healthcare settings in Bangladesh, improving transportation accessibility, waiting-area conditions, consistent availability of free medicines, and patient-centered communication among healthcare staff may substantially enhance patient experience and quality of care.

Introduction

Patient satisfaction is widely regarded as a central indicator of healthcare quality and system performance. Beyond reflecting patients’ subjective experiences, it captures important dimensions of care such as adherence to medical advice, effective utilization of services, the quality of the doctor–patient relationship, and continuity and coordination of care [1]. As a multidimensional construct, patient satisfaction provides insight into both clinical effectiveness and the responsiveness of health services to patient needs. Higher levels of satisfaction are associated with improved treatment adherence, greater patient retention, reduced malpractice risk, and more efficient and patient-centered delivery of care [2, 3]. Consequently, understanding the determinants of patient satisfaction is critical for strengthening healthcare systems, particularly in resource-constrained settings.

Evidence indicates that patient satisfaction is shaped by a combination of socio-demographic characteristics, interpersonal processes, and structural features of healthcare delivery. Socio-demographic factors such as age, education, and socioeconomic position have been consistently associated with variation in satisfaction levels, with older individuals and those with lower educational attainment often reporting higher satisfaction [4]. However, interpersonal aspects of care frequently exert a stronger influence. Effective communication, respectful treatment, responsiveness to patient concerns, and fulfillment of expectations during clinical encounters are key drivers of satisfaction [5, 6]. In addition, organizational and system-level characteristics, including staffing adequacy, waiting time, affordability of services, facility cleanliness, and alignment between perceived and delivered quality, substantially shape patient experiences across diverse healthcare contexts [7–10]. Across both high-income and low- and middle-income countries, unmet care needs, poor communication, and insufficient follow-up are consistently associated with lower satisfaction, whereas positive service experiences increase patients’ willingness to return for future care [6].

Similar determinants of patient satisfaction have been documented across Asian and South Asian healthcare systems, where interpersonal and structural dimensions of care jointly shape patient experiences [11–14]. Evidence from the region consistently reports the importance of respectful, responsive, and empathetic provider–patient interactions in fostering trust and enhancing overall satisfaction [11, 12]. Communication quality, particularly clarity of explanations, opportunities for patient engagement, and adequate consultation time, has emerged as a critical determinant in both primary and secondary care settings [11, 12]. Concurrently, facility-level factors such as cleanliness, waiting time, availability of essential medicines, and affordability of services substantially influence perceptions of care quality [13, 14]. Variations in satisfaction across age, income, and educational groups further suggest that patient expectations and experiences are shaped by broader social and structural inequalities within regional health systems [12, 14]. This body of evidence suggests that patient satisfaction extends beyond clinical competence to encompass the broader organization, accessibility, and responsiveness of healthcare delivery.

Upazila Health Complexes represent the first referral tier of Bangladesh’s public health system, providing outpatient, maternal, child, and basic inpatient services to both rural and peri-urban populations, and are often the primary point of contact with formal healthcare for low-income households [15]. Understanding patient satisfaction at this level is therefore critical for strengthening primary care responsiveness and equity in service delivery. In Bangladesh, existing evidence suggests that patient satisfaction within public healthcare facilities remains moderate, with reported levels typically ranging between about 50–65% of service users, depending on facility type and service context [15, 16]. In a government-financed primary healthcare setting in Bangladesh, 55.1% of patients reported being satisfied with services [16]. Satisfaction was positively associated with provider behavior, privacy, and registration efficiency, while longer waiting times and administrative delays were linked to lower perceived service quality. Availability of essential medications and access to diagnostic services have also been identified as important contributors to higher satisfaction in public facilities [16]. Research at the primary care level further highlights the influence of patient-centered physician behaviors (including effective communication, nonverbal engagement, and adequate consultation time) on overall satisfaction with medical consultations [17]. In addition, facility-level characteristics such as cleanliness, privacy arrangements, operating hours, and physical accessibility continue to shape patient experiences across both public and private settings, with satisfaction often reported as lower in primary care facilities compared to private institutions [15]. At the upazila level, provider attitude and responsiveness appear to exert the strongest influence on patient satisfaction, while infrastructural resources and accessibility factors contribute in a moderate yet meaningful manner to perceived service quality [18]. Although the present study is based on a single facility, the selected Upazila Health Complex reflects typical characteristics of public primary healthcare settings in Bangladesh, making it a relevant context for examining patient satisfaction.

Therefore, the present study aimed to examine the determinants of patient satisfaction at a primary-level public healthcare facility in Bangladesh. While previous studies have primarily relied on conventional statistical approaches at tertiary and district-level facilities, no study to our knowledge has applied predictive machine-learning models to data from primary-level upazila health complexes. Additionally, there is limited facility-level evidence examining the relative importance of interpersonal, environmental, and access-related determinants simultaneously. Although regression-based approaches are widely used and interpretable, they assume linear relationships and require pre-specification of interaction terms, which may limit their ability to capture complex and non-linear interactions among predictors such as staff behavior, waiting time, and access to services. In contrast, machine-learning approaches, particularly ensemble tree-based models such as Random Forest and XGBoost, can identify non-linear relationships and higher-order interactions without prior specification, making them well-suited for patient satisfaction data where multiple dimensions of care interact. In this study, machine learning is used alongside traditional regression to complement explanatory analysis and enhance predictive performance. From a practical perspective, predictive models can support facility-level decision-making by identifying patient groups at higher risk of dissatisfaction, enabling targeted and data-driven service improvements.

Methods

Study participants and procedure

This cross-sectional study was conducted at the Savar Upazila Health Complex, a government-run primary-level public healthcare facility located within Savar Municipality of Dhaka District, Bangladesh. Upazila Health Complexes in Bangladesh primarily deliver outpatient, maternal, child, preventive, and basic inpatient services and function as the first referral level for public healthcare system. Although limited inpatient services are available, the facility does not operate as a tertiary or specialized hospital. The facility was purposively selected because it serves a mixed urban–rural population and represents a typical government-run Upazila Health Complex serving a large catchment population. As such, it reflects common characteristics of primary healthcare facilities in Bangladesh, making it a relevant setting for examining patient satisfaction in similar contexts.

The study targeted individuals who accessed medical services at the Savar Upazila Health Complex during the data collection period from 15 April to 8 May 2025. Eligible participants were patients aged 10 years and above who received outpatient or inpatient services during the study period and were capable of providing informed consent or assent; for minors (< 18 years), written informed consent was obtained from a parent or legal guardian and assent was obtained from the participant. The age threshold of 10 years was selected to ensure that participants had sufficient cognitive ability to understand and respond meaningfully to questions related to their healthcare satisfaction experiences [19, 20].

Interviews with minor participants were conducted directly with the participants, while parents or guardians were present but did not participate in answering the questions, ensuring that responses reflected the participants’ own perceptions and experiences. For complex or factual variables (e.g., household income, distance to the hospital, duration of illness, and waiting time), minor participants were allowed to consult their accompanying parents or guardians. In such cases, parents or guardians provided clarification only for objective information, while all perception-based responses, including satisfaction-related questions, were answered solely by the participants.

Exclusion criteria included patients in critical condition, those with cognitive impairment or severe illness preventing participation, and individuals unwilling or unable to complete the interview. A convenience sampling strategy was employed due to logistical and resource constraints typical of facility-based studies in low-resource settings. Trained data collectors approached eligible patients after completion of their consultation or treatment in outpatient waiting areas or inpatient wards.

All interviews were conducted face-to-face using a structured questionnaire administered via mobile devices. The study objectives were explained prior to participation, and written informed consent or assent was obtained. To minimize selection bias, data were collected across multiple time periods (morning, noon, and afternoon shifts) and on all days of the week, ensuring inclusion of patients attending at different times and reducing potential temporal selection bias. During each shift, all eligible patients present in waiting areas and wards were approached for participation. Interviews were conducted in a quiet area within the facility to ensure privacy. In total, 508 participants were included in the study.

Measures

Sociodemographic factors

Sociodemographic characteristics of the participants were collected to capture their demographic and socioeconomic background. The variables included gender (male or female), age category (10–25, 26–50, or ≥ 51 years), current residence (rural or urban), family income level (lower, middle, or higher), and educational attainment (no formal education, up to primary, secondary/higher secondary, or honors and above). Additional variables included occupation (unemployed, employed, student, housewife, or business owner), distance from home to the hospital (< 5 km or ≥ 5 km), and convenience of transportation to the hospital (bad, moderate, or good).

Patient related factors

Patient-related health variables included the primary disease category (presenting condition) for which the patient sought care (digestive, dermatological, reproductive, injury, fever, rheumatological, respiratory, dental, endocrine, and others) and duration of illness (7 days or less, 8–14 days, 15–30 days, and more than 30 days). The primary disease category represents the main health complaint for the current visit. Admission status (inpatient vs. outpatient) was recorded to indicate whether the patient was hospitalized at the time of data collection. Presence of chronic disease (yes or no) was also recorded to capture underlying long-term health conditions, such as diabetes, hypertension, or other chronic illnesses.

Hospital and facilities related variables

Hospital and facility-related variables were collected to assess patients’ experiences with service delivery, facility conditions, and staff responsiveness during their visit to the health facility. These variables were based on patients’ self-reported perceptions at the time of the interview. Service-related factors included waiting time before receiving care (less than 15 min, 15–30 min, 30–60 min, or more than 60 min) and receipt of free medicine (yes or no). Facility conditions were assessed through the availability of comfortable seating in the waiting area, tranquil and undisturbed surroundings, access to healthy food, and access to healthy toilet facilities (each recorded as yes or no), as well as patients’ perception of the overall hospital environment (bad, moderate, or good). Patients’ perceptions of staff behavior were also evaluated, including the behavior of the receptionist and behavior of nurses (unhelpful, moderately helpful, or highly helpful). In addition, privacy of the prescription during consultation (yes or no) was recorded to assess confidentiality during patient care.

Patient satisfaction (PSQ-18)

Patient satisfaction was measured using the Patient Satisfaction Questionnaire Short-Form (PSQ-18), a validated instrument designed to assess multiple dimensions of satisfaction with medical care [21]. The questionnaire was translated into Bangla using a forward–backward translation procedure to ensure semantic equivalence with the original instrument. The translated version was reviewed by public health experts and pretested among a small group of patients prior to data collection. A formal psychometric validation of the translated instrument was not conducted in this study and is acknowledged as a limitation.

The internal consistency reliability of the scale in the present study was acceptable, with a Cronbach’s alpha coefficient of 0.83. The PSQ-18 consists of 18 items grouped into seven subscales: General Satisfaction (items 3 and 17), Technical Quality (items 2, 4, 6, and 14), Interpersonal Manner (items 10 and 11), Communication (items 1 and 13), Financial Aspects (items 5 and 7), Time Spent with Doctor (items 12 and 15), and Accessibility and Convenience (items 8, 9, 16, and 18). Each item is rated on a five-point Likert scale ranging from 1 (strongly agree) to 5 (strongly disagree), with specific items reverse-coded so that higher scores consistently indicate greater satisfaction.

The total satisfaction score was calculated by summing the responses across all 18 items, resulting in a possible score range of 18–90. To categorize satisfaction levels, the percentage of the possible score range was calculated using the following formula: Adjusted percentage = (score − 18) / (90 − 18) ×100. Based on Bloom’s cutoff, scores below 61 (< 60% of the possible score range above the minimum) were classified as poor satisfaction, whereas scores of 61 or higher were categorized as moderate to good satisfaction. This has been used in a previous study [22]. The PSQ-18 demonstrated good internal consistency reliability in the present study, with a Cronbach’s alpha coefficient of 0.830.

Ethical consideration

The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki (2013). Ethical approval was granted by the Thesis Proposal Approval Committee (2024), Department of Government and Politics, Jahangirnagar University, following evaluation of the research proposal on 4 December 2024. Written informed consent was obtained from all participants prior to data collection. For participants aged 10 to below 18 years, parental or guardian consent was obtained in accordance with ethical requirements. Permission to conduct the study was also obtained from the hospital authorities. Participants were informed about the study’s purpose, procedures, and their right to withdraw at any time without consequences. Participation in the study was voluntary, and no incentives were provided. To ensure confidentiality, all collected data were anonymized and securely stored throughout the study, and the data were used solely for research purposes.

Statistical analysis

Following data collection, responses were initially recorded in Google Forms and subsequently exported to Microsoft Excel 2024 for data cleaning and preparation. The cleaned dataset was then analyzed using the Statistical Package for the Social Sciences (SPSS), version 27. Descriptive statistics were computed to summarize participants’ characteristics and study variables, and were presented as frequencies, percentages, means, and standard deviations. To examine preliminary associations between patient satisfaction scores and independent variables, independent sample t-tests and one-way analysis of variance (ANOVA) were conducted, as appropriate. Variables that demonstrated statistical significance in univariable analyses (independent t-test or one-way ANOVA), along with variables considered theoretically or contextually relevant based on prior literature, were entered into a multivariable binary logistic regression model to identify factors associated with higher patient satisfaction. The dependent variable, patient satisfaction was categorized as poor satisfaction and moderate to good satisfaction based on the PSQ-18 scoring criteria, following Bloom's cutoff. Adjusted odds ratios (AORs) with corresponding 95% confidence intervals (CIs) were reported to estimate the strength of associations. Model adequacy was evaluated using the Hosmer–Lemeshow goodness-of-fit test, with a non-significant result indicating acceptable model fit. Statistical significance for all inferential tests was set at p < 0.05. Overall, both logistic regression and machine-learning models were fitted using the same set of predictor variables to ensure consistency and comparability. Logistic regression was used to identify significant associations between predictors and patient satisfaction, while machine-learning models were applied to enhance predictive performance and capture potential non-linear relationships.

Machine learning analysis

A supervised classification pipeline was developed to predict overall patient satisfaction. Analyses were conducted in PyCharm on a workstation equipped with an NVIDIA GeForce RTX 4060 GPU. The dataset was divided using a stratified 70:30 split, resulting in 355 training instances and 153 test instances. To address class imbalance while avoiding information leakage, the Synthetic Minority Over-sampling Technique (SMOTE) was applied only within the training folds. Feature scaling was performed using StandardScaler. Six classifiers were evaluated: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). GPU acceleration was enabled where supported (XGBoost tree_method=gpu_hist, LightGBM device = gpu, CatBoost task_type = GPU). All models were implemented within an imbalanced-learn pipeline to ensure consistent preprocessing across classifiers.

Hyperparameter optimization was performed via grid search with 5-fold stratified cross-validation (shuffle, random_state = 42) on the training set. To prevent leakage, SMOTE and StandardScaler were applied within each training fold. The primary optimization objective was ROC-AUC to evaluate discrimination across thresholds, while additional metrics including accuracy, precision, recall and F1-score were also recorded for transparency. Because practical deployment may prioritize identifying dissatisfied patients, recall and F1-score were additionally considered to reflect the trade-off between discrimination and error balance. After tuning, model performance was evaluated on the held-out test set using a panel of metrics including accuracy, ROC-AUC, precision, recall (sensitivity), F1-score, and log loss. Confusion matrices were generated for each model, and receiver operating characteristic (ROC) curves were plotted to enable visual comparison. For the primary model, global feature importance was computed from the tree-based ensemble using Gini importance.

Model preference was determined according to the analytic objective. ROC-AUC and log loss were used to assess overall discrimination and probability calibration across thresholds, while recall (sensitivity) was prioritized when minimizing false negatives was considered important (i.e., failing to identify dissatisfied patients). Because class imbalance can inflate accuracy, it was interpreted cautiously alongside precision, recall, and F1-score. When models demonstrated similar ROC-AUC values, preference was given to those with higher recall and F1-score and lower log loss, indicating a better balance between classification performance and reliable probability estimates. To aid interpretation, confusion matrices were reported at a prespecified operating threshold and ROC curves were presented to illustrate discrimination–recall trade-offs.

Results

Description of the study participants

Most participants were female (69.9%), aged 26–50 years (49.0%), and resided in urban areas (64.0%). The majority were from lower-income households (56.1%). Secondary or higher secondary education was the most common educational level (36.8%), and housewives comprised the largest occupational group (46.9%). Regarding healthcare access, 41.9% of participants reported good transportation convenience and 55.5% lived within 5 km of the facility. Reproductive system–related conditions were the most common presenting problems (15.0%). Most participants reported symptom duration of ≤ 7 days (53.7%), were not admitted (82.9%), and did not report chronic disease (72.8%). The most common waiting time was 15–30 min (31.7%).

In terms of facility-related experiences, 46.9% rated the hospital environment as moderate. Most participants reported comfortable seating (75.0%), tranquil surroundings (70.5%), availability of healthy food (65.7%), and clean toilet facilities (58.1%). Receptionist and nurse behavior were rated as helpful by 53.9% and 58.1% of participants, respectively, and 84.1% reported receiving free medicines (Table 1).

Table 1.

Bivariate associations between independent variables and patient satisfaction at an Upazila Health Complex in Bangladesh

Variables	(n), %	Patient Satisfaction Score (Mean ± SD)	Test statistic (t/F) and p-value
Sociodemographic Variables
Gender
Male	153 (30.1)	63.67 ± 8.03	− 0.91 (0.36)
Female	355 (69.9)	64.47 ± 9.19
Age group
10 to 25	171 (33.7)	65.36 ± 8.84	3.01 (0.05)
26 to 50	249 (49)	64.05 ± 8.71
51 or more	88 (17.3)	62.57 ± 9.09
Current residence
Rural	183 (36)	63.60 ± 7.75	− 1.22 (0.224)
Urban	325 (64)	64.59 ± 9.42
Family income
Lower income	285 (56.1)	63.93 ± 7.42	0.51 (0.60)
Middle income	131 (25.8)	64.88 ± 9.65
High income	92 (18.1)	64.23 ± 11.45
Educational attainment
No education	112 (22)	64.00 ± 8.48	0.66 (0.55)
Up to primary	170 (33.5)	64.77 ± 9.96
SSC/HSC	187 (36.8)	64.23 ± 7.50
Honors or more	39 (7.7)	62.56 ± 10.76
Occupation
Unemployed	74 (14.6)	62.89 ± 10.08	1.004 (0.41)
Employed	86 (16.9)	65.00 ± 7.78
Student	72 (14.2)	64.99 ± 8.49
Housewife	238 (46.9)	64.38 ± 9.13
Business	38 (7.5)	62.74 ± 7.29
Convenient transportation to hospital
Bad	96 (18.9)	61.33 ± 9.00	19.86 (< 0.001)
Moderate	199 (39.2)	62.68 ± 8.77
Good	213 (41.9)	66.99 ± 8.12
Distance to hospital from home
Less than 5 KM	282 (55.5)	64.39 ± 7.90	0.44 (0.66)
More than 5 KM	226 (44.5)	64.04 ± 9.93
Patient Related Information
Patients’ disease related systems
Digestive System	63 (12.4)	63.79 ± 10.00	1.23 (0.28)
Dermatological System	60 (11.8)	63.48 ± 8.18
Reproductive System	76 (15)	66.14 ± 7.48
Injury	35 (6.9)	62.71 ± 8.61
Fever	42 (8.3)	64.93 ± 9.37
Rheumatology	67 (13.2)	63.01 ± 8.75
Respiratory System	47 (9.3)	62.57 ± 7.26
Others	62 (12.2)	64.63 ± 11.32
Dental	32 (6.3)	66.72 ± 9.01
Endocrine	24 (4.7)	64.50 ± 5.20
Days of being affected to a disease
7 days or less	273 (53.7)	64.69 ± 8.84	1.23 (0.30)
8 to 14 days	33 (6.5)	65.64 ± 11.92
15 to 30 days	46 (9.1)	62.76 ± 7.91
More than 30 days	156 (30.7)	63.57 ± 8.38
Admitted to the hospital
No	421 (82.9)	63.62 ± 8.58	− 3.48 (< 0.001)
Yes	87 (17.1)	67.21 ± 9.61
Presence of chronic disease
No	370 (72.8)	64.32 ± 8.63	0.35 (0.73)
Yes	138 (27.2)	64.01 ± 9.47
Hospital and facilities related variables
Waiting time
More than 60 min	88 (17.3)	61.90 ± 11.23	8.93 (< 0.001)
31 to 60 min	135 (26.6)	63.76 ± 8.39
16 to 30 min	161 (31.7)	63.34 ± 8.30
Less than 15 min	124 (24.4)	67.56 ± 7.18
Hospital environment
Bad	58 (11.4)	55.74 ± 12.77	51.84 (< 0.001)
Moderate	238 (46.9)	63.31 ± 7.29
Good	212 (41.7)	67.59 ± 7.25
Comfortable waiting room
No	127 (25)	58.99 ± 11.30	− 8.19 (< 0.001)
Yes	381 (75)	65.98 ± 7.08
Tranquil and undisturbed surroundings
No	150 (29.5)	60.33 ± 10.82	− 6.71 (< 0.001)
Yes	358 (70.5)	65.87 ± 7.31
Availability of healthy food
No	174 (34.3)	59.63 ± 10.48	− 9.12 (< 0.001)
Yes	334 (65.7)	66.63 ± 6.75
Availability of healthy toilet
No	213 (41.9)	60.92 ± 10.23	− 7.55 (< 0.001)
Yes	295 (58.1)	66.62 ± 6.79
Behavior of the receptionist
Unhelpful	56 (11)	56.44 ± 11.98	76.12 (< 0.001)
Moderate	178 (35.1)	60.91 ± 7.44
Helpful	274 (53.9)	67.98 ± 6.87
Behavior of the nurses
Unhelpful	28 (5.5)	54.64 ± 14.33	77.22 (< 0.001)
Moderate	185 (36.4)	60.08 ± 7.68
Helpful	295 (58.1)	67.75 ± 6.89
Receipt of free medicine
No	81 (15.9)	58.01 ± 11.42	− 7.24 (< 0.001)
Yes	427 (84.1)	65.41 ± 7.75
Privacy of the prescription
No	208 (40.9)	64.14 ± 8.17	0.02 (0.98)
Yes	300 (59.1)	64.29 ± 9.32

Bivariate associations between patient satisfaction and study variables

The overall mean patient satisfaction score was 64.23 ± 8.86. Among sociodemographic variables, age group showed a marginal association with satisfaction (F = 3.01, p = 0.05), with younger participants reporting slightly higher satisfaction. Transportation convenience was strongly associated with satisfaction (F = 19.86, p < 0.001) and hospital admission status was also associated with satisfaction (t = -3.48, p < 0.001).

Several facility-related factors showed significant associations with patient satisfaction, including waiting time, hospital environment, comfortable waiting areas, tranquil surroundings, availability of healthy food, and toilet facilities (all p < 0.001). Staff behavior, including receptionist and nurse behavior, was also strongly associated with satisfaction (Table 1).

Factors associated with patient satisfaction

Table 2 presents the results of the multivariable binary logistic regression analysis examining factors associated with higher patient satisfaction. The model demonstrated moderate explanatory power (pseudo-R² = 0.444). The Hosmer–Lemeshow goodness-of-fit test indicated an adequate model fit (χ² = 6.385, p = 0.604).

Table 2.

Multivariable logistic regression analysis of factors associated with higher patient satisfaction at an Upazila Health Complex in Bangladesh

Name of the Variables	B	S.E.	Sig.	Exp(B)	pseudo-R² = 0.444
					95% C. I. for EXP(B)
					Lower	Upper
Gender (Female) [Ref: Male]	0.127	0.416	0.761	1.135	0.502	2.567
Age category [Ref: 51 or more]			0.650
10 to 25	-0.147	0.535	0.784	0.864	0.302	2.466
26 to 50	-0.339	0.420	0.420	0.712	0.313	1.623
Current residence (Urban) [Ref: Rural]	-0.176	0.344	0.608	0.838	0.427	1.645
Family income [Ref: Lower income]			0.762
Middle income	-0.243	0.346	0.482	0.784	0.398	1.545
High income	-0.019	0.401	0.963	0.982	0.448	2.152
Educational qualification [Ref: Honors or more]			0.544
No education	0.842	0.646	0.192	2.321	0.655	8.226
Up to primary	0.444	0.560	0.428	1.559	0.520	4.674
SSC/HSC	0.637	0.549	0.246	1.891	0.645	5.548
Occupation [Ref: Business]			0.411
Unemployed	-0.150	0.654	0.818	0.860	0.239	3.097
Employed	0.549	0.604	0.363	1.731	0.530	5.653
Student	-0.099	0.693	0.886	0.906	0.233	3.520
Housewife	-0.413	0.660	0.531	0.661	0.181	2.414
Convenient transportation to hospital [Ref: Bad]			0.119
Moderate	0.479	0.389	0.218	1.615	0.754	3.461
Good	0.933	0.454	0.040	2.542	1.044	6.190
Distance to hospital (More than 5 KM) [Ref: Less than 5 KM]	0.083	0.307	0.786	1.087	0.596	1.984
Patients’ disease related systems [Ref: Digestive System]			0.603
Dermatological System	0.898	0.569	0.114	2.455	0.806	7.483
Reproductive System	0.841	0.540	0.119	2.320	0.805	6.689
Injury	0.454	0.630	0.471	1.574	0.458	5.407
Fever	0.710	0.602	0.238	2.034	0.625	6.615
Rheumatology	0.624	0.525	0.235	1.867	0.667	5.228
Respiratory System	0.530	0.551	0.336	1.699	0.577	5.002
Others	0.138	0.523	0.792	1.148	0.412	3.200
Dental	1.654	0.783	0.035	5.227	1.126	24.255
Endocrine	0.853	0.801	0.287	2.348	0.489	11.279
Days of being affected to a disease [Ref: More than 30 days]			0.415
7 days or less	0.574	0.357	0.108	1.776	0.882	3.576
8 to 14 days	0.001	0.589	0.998	1.001	0.315	3.178
15 to 30 days	0.182	0.519	0.725	1.200	0.434	3.319
Admitted to the hospital (No) [Ref: Yes]	0.119	0.444	0.789	1.126	0.472	2.689
Presence of chronic disease (Yes) [Ref: No]	0.174	0.365	0.633	1.190	0.582	2.432
Waiting time [Ref: More than 60 min]			0.233
30 to 60 min	0.259	0.436	0.553	1.296	0.551	3.048
15 to 30 min	0.724	0.423	0.087	2.062	0.901	4.722
Less than 15 min	0.754	0.476	0.113	2.125	0.836	5.403
Hospital Environment [Ref: Bad]			0.933
Moderate	0.151	0.433	0.728	1.162	0.497	2.716
Good	0.185	0.534	0.729	1.203	0.422	3.428
Comfortable waiting room (Yes) [Ref: No]	0.777	0.341	0.023	2.175	1.116	4.241
Tranquil and undisturbed surroundings (Yes) [Ref: No]	-0.122	0.352	0.728	0.885	0.443	1.766
Availability of healthy food (Yes) [Ref: No]	0.908	0.350	0.009	2.480	1.249	4.925
Availability of healthy toilet (Yes) [Ref: No]	0.399	0.332	0.229	1.490	0.778	2.853
Receipt of free medicine (Yes) [Ref: No]	0.854	0.352	0.015	2.349	1.178	4.686
Behavior of the receptionist [Ref: Unhelpful]			0.413
Moderate	0.031	0.506	0.951	1.031	0.382	2.783
Helpful	0.573	0.608	0.346	1.774	0.538	5.844
Behavior of the nurses [Ref: Unhelpful]			< 0.001
Moderate	-1.117	0.708	0.115	0.327	0.082	1.313
Helpful	0.366	0.767	0.633	1.442	0.321	6.479
Privacy of the prescription (Yes) [Ref: No]	-0.260	0.298	0.383	0.771	0.430	1.383
Constant	-2.927	1.244	0.019	0.054

Transportation convenience remained a significant predictor of satisfaction. Patients reporting good transportation access had higher odds of satisfaction compared with those reporting poor access (AOR = 2.54, 95% CI: 1.04–6.19, p = 0.040). Among disease categories, patients presenting with dental conditions were more likely to report higher satisfaction compared with those with digestive conditions (AOR = 5.23, 95% CI: 1.13–24.26, p = 0.035).

Several facility-related factors were also significantly associated with satisfaction, including comfortable seating in the waiting area (AOR = 2.18, 95% CI: 1.12–4.24, p = 0.023), availability of healthy food (AOR = 2.48, 95% CI: 1.25–4.93, p = 0.009), and receipt of free medicines (AOR = 2.35, 95% CI: 1.18–4.69, p = 0.015). Nurse behavior showed a significant overall association with patient satisfaction (p < 0.001), indicating its joint effect. However, individual category comparisons relative to the reference group were not statistically significant.

Machine learning model performance

Table 3 summarizes the baseline test-set performance of the candidate models prior to hyperparameter tuning. Among the models, Random Forest (RF) demonstrated the strongest baseline performance, achieving the highest accuracy (0.7843), the highest area under the receiver operating characteristic curve (AUC = 0.7651), and the lowest log loss (0.477). CatBoost achieved the highest recall (0.9211), while XGBoost and LightGBM produced the highest precision values (0.8226). The remaining models showed comparable performance across accuracy, F1-score, and log-loss metrics.

Table 3.

Test-set performance of machine learning models before hyperparameter tuning

Model	Accuracy	AUC	Precision	Recall	F1-Score	Log Loss
LR	0.7516	0.7643	0.8167	0.8596	0.8376	0.5217
RF	0.7843	0.7651	0.8189	0.9123	0.8631	0.477
XGBoost	0.7778	0.6853	0.8226	0.8947	0.8571	0.8364
LightGBM	0.7778	0.7024	0.8226	0.8947	0.8571	0.6866
CatBoost	0.7778	0.7506	0.8077	0.9211	0.8607	0.5321
SVM	0.7778	0.7542	0.8125	0.9123	0.8595	0.5004

The cross-validation performance after hyperparameter optimization is presented in Table 4. Mean cross-validated AUC values ranged from 0.7214 to 0.7945, indicating broadly similar discrimination across models. Logistic Regression (LR) achieved the highest mean AUC (0.7945 ± 0.0610), while Random Forest produced the highest mean accuracy (0.7972 ± 0.0303), recall (0.8707 ± 0.0280), and F1-score (0.8640 ± 0.0213). Support Vector Machine (SVM) also demonstrated competitive performance, with a mean F1-score of 0.8555 ± 0.0346 and recall of 0.8667 ± 0.0409. The gradient-boosting models (XGBoost, LightGBM, and CatBoost) showed comparable performance but did not exceed RF or LR across the evaluated metrics.

Table 4.

Cross-validated performance of machine learning models (5-fold stratified CV; mean ± SD) and selected hyperparameters

Model	AUC	Accuracy	Precision	Recall	F1-Score	Best Params
LR	0.7945 ± 0.0610	0.7775 ± 0.0483	0.8576 ± 0.0436	0.8401 ± 0.0363	0.8483 ± 0.0339	{‘classifier__C’: 0.1, ‘classifier__penalty’: ‘l1’}
RF	0.7828 ± 0.0627	0.7972 ± 0.0303	0.8581 ± 0.0259	0.8707 ± 0.0280	0.8640 ± 0.0213	{‘classifier__max_depth’: 5, ‘classifier__n_estimators’: 100}
XGBoost	0.7473 ± 0.0641	0.7690 ± 0.0363	0.8460 ± 0.0413	0.8440 ± 0.0307	0.8441 ± 0.0239	{‘classifier__learning_rate’: 0.05, ‘classifier__max_depth’: 3, ‘classifier__n_estimators’: 100, ‘classifier__tree_method’: ‘gpu_hist’}
LightGBM	0.7214 ± 0.0551	0.7493 ± 0.0225	0.8289 ± 0.0345	0.8366 ± 0.0251	0.8319 ± 0.0133	{‘classifier__device’: ‘gpu’, ‘classifier__learning_rate’: 0.05, ‘classifier__n_estimators’: 100, ‘classifier__num_leaves’: 20}
CatBoost	0.7557 ± 0.0680	0.7662 ± 0.0374	0.8312 ± 0.0446	0.8631 ± 0.0368	0.8455 ± 0.0238	{‘classifier__depth’: 6, ‘classifier__iterations’: 200, ‘classifier__learning_rate’: 0.05, ‘classifier__task_type’: ‘GPU’}
SVM	0.7855 ± 0.0595	0.7831 ± 0.0509	0.8462 ± 0.0468	0.8667 ± 0.0409	0.8555 ± 0.0346	{‘classifier__C’: 0.1, ‘classifier__gamma’: ‘scale’}

Following hyperparameter tuning, model performance was evaluated on the held-out test set (Table 5). XGBoost achieved the highest overall classification performance, with the highest accuracy (0.7974), recall (0.9211), and F1-score (0.8714). Random Forest demonstrated the strongest discrimination and calibration, achieving the highest AUC (0.8198) and the lowest log loss (0.4534). Logistic Regression achieved the highest precision (0.8306) among the evaluated models. The remaining models showed comparable but slightly lower performance across the evaluated metrics.

Table 5.

Test-set performance of tuned machine learning models

Model	Accuracy	AUC	Precision	Recall	F1-Score	Log Loss
LR	0.7908	0.7994	0.8306	0.9035	0.8655	0.4715
RF	0.7908	0.8198	0.8254	0.9123	0.8667	0.4534
XGBoost	0.7974	0.8045	0.8268	0.9211	0.8714	0.4687
LightGBM	0.7843	0.7490	0.8140	0.9211	0.8642	0.5571
CatBoost	0.7908	0.7355	0.8254	0.9123	0.8667	0.5932
SVM	0.7778	0.7978	0.8175	0.9035	0.8583	0.4861

Error profiles from confusion matrices

Figure 1 presents the confusion matrices for the tuned classifiers on the held-out test dataset. Across models, the number of correctly identified satisfied patients (true positives) ranged from 103 to 105 out of 114, whereas correctly identifying dissatisfied patients (true negatives) ranged from 15 to 18 out of 39. Logistic Regression produced the highest number of true negatives (TN = 18). Random Forest and CatBoost produced identical classification outcomes (TN = 17, FP = 22, FN = 10, TP = 104). XGBoost produced the lowest number of false negatives (FN = 9) and the highest number of true positives (TP = 105). LightGBM also produced nine false negatives but fewer true negatives (TN = 15). SVM produced 11 false negatives and 23 false positives.

Fig. 1.

Confusion matrices for tuned classifiers on the test set. Panels show (A) Logistic Regression, (B) Random Forest, (C) XGBoost, (D) LightGBM, (E) CatBoost, and (F) Support Vector Machine. Rows represent true labels and columns represent predicted labels

Receiver operating characteristic (ROC) comparison

Figure 2(A) shows the receiver operating characteristic (ROC) curves for the tuned models on the test dataset. Random Forest achieved the highest AUC (0.8198), followed by XGBoost (0.8045), Logistic Regression (0.7994), and SVM (0.7978). LightGBM (0.7490) and CatBoost (0.7355) demonstrated comparatively lower discrimination performance. Figure 2(B) presents the relationship between ROC AUC and log loss across models. Random Forest achieved both the highest AUC and the lowest log loss (0.4534), indicating the strongest combined performance among the evaluated classifiers. XGBoost (log loss = 0.4687) and Logistic Regression (log loss = 0.4715) showed similar AUC values with slightly higher log loss. LightGBM and CatBoost showed lower AUC values and higher log loss compared with the other models.

Fig. 2.

Performance comparison of tuned models on the test set. (A) Receiver operating characteristic (ROC) curves. (B) Relationship between ROC AUC and log loss across models

Global feature importance

Figure 3 presents the global feature importance derived from the tuned Random Forest model based on the mean decrease in Gini impurity. The most influential predictors of patient satisfaction were behavior of the receptionist and comfortable waiting room conditions, followed by behavior of nurses, availability of healthy food, and availability of healthy toilets. Other variables with notable importance included receipt of free medicine, hospital environment, transportation convenience, and waiting time. In contrast, sociodemographic variables such as occupation, educational qualification, family income, and age category showed comparatively lower importance values. Variables including current residence, prescription privacy, distance to the hospital, and gender demonstrated minimal contribution to the model.

Fig. 3.

Global feature importance derived from the tuned Random Forest model based on mean decrease in Gini impurity

Discussion

This study examined the determinants of patient satisfaction at a primary healthcare facility in Bangladesh, while also exploring the use of machine-learning approaches as an exploratory tool to better understand complex relationships among satisfaction determinants. Overall, the findings indicate a moderate level of patient satisfaction. The results suggest that multiple dimensions of care influence patient satisfaction, particularly factors related to accessibility, facility conditions, and interpersonal interactions with healthcare staff. In particular, transportation convenience, comfortable waiting areas, access to healthy food, and the receipt of free medicines were associated with higher levels of satisfaction. Interpersonal aspects of care, especially nursing staff behavior, also emerged as an important determinant. These findings highlight that both structural elements of healthcare delivery and the quality of patient–provider interactions play a central role in shaping patients’ experiences in primary healthcare settings.

This study found a mean patient satisfaction score of 64.23 ± 8.86 among individuals attending the Savar Upazila Health Complex. Considering that the total PSQ-18 score ranges from 18 to 90, this value indicates a moderate level of overall satisfaction among the study participants. This level of satisfaction appears broadly consistent with findings from previous studies that used the same instrument in different healthcare contexts. For example, a study conducted in a primary healthcare setting in Lithuania reported a mean satisfaction score of 59.9 (± 14.6) [23], while study from Malaysia using the PSQ-18 documented a mean score of 59.2 (± 6.5) [11]. Similarly, a study from the United States assessing satisfaction with medical care among patients with chronic low back pain reported a comparable mean score of 62.6 (± 28.1) [24]. In Bangladesh, however, a recent study conducted at the Government Unani & Ayurvedic Medical College Hospital in Dhaka reported a somewhat higher mean PSQ-18 score of 69.1 (± 16.9) [25]. The relatively lower score observed at the Savar Upazila Health Complex may be partly explained by contextual differences between the healthcare facilities. Savar operates as a high-volume primary healthcare center serving both urban residents and surrounding rural populations, where higher patient load, longer travel distances, and greater service demand may influence patients’ perceptions of convenience and overall care experience. In addition, the inclusion of both inpatient and outpatient participants and the relatively larger sample size in the present study may have contributed to greater variability in satisfaction responses. Despite these contextual differences, the overall level of patient satisfaction observed in this study remains broadly comparable with findings reported from other national and international healthcare settings using the PSQ-18 instrument.

The present study identified transportation convenience to the hospital as a significant determinant of higher patient satisfaction. Patients who reported good transportation access had approximately 2.5 times higher odds of satisfaction compared with those reporting poor transportation access. This finding highlights the important role of physical accessibility in shaping patients’ experiences within primary healthcare settings. Convenient transportation may facilitate timely healthcare utilization and reduce barriers to seeking care when needed. Previous evidence supports this observation. For instance, a study in the United States reported that approximately 5.8 million people (1.8%) delayed medical care due to transportation barriers in 2017 [26]. Similarly, healthcare access frameworks recognize transportation as a key enabling factor influencing healthcare utilization and patient experience [27]. A systematic review also identified transportation barriers as an important obstacle to healthcare access, particularly among individuals with lower incomes or those without insurance coverage [28]. Transportation barriers may disproportionately affect individuals who require frequent healthcare visits, including people with disabilities, patients with chronic diseases, and pregnant women [27, 29, 30]. Studies from other contexts further support this relationship; for example, transportation difficulties have been associated with unmet healthcare needs among rural elderly populations in Korea [31], while travel time to healthcare facilities has been shown to significantly influence patient satisfaction in Nepal [14]. In Bangladesh, limited or unreliable transportation has been reported as a major barrier to accessing maternal healthcare services during pregnancy, delivery, and the postpartum period [30]. These findings collectively suggest that improving transportation accessibility may play an important role in enhancing healthcare utilization and overall patient satisfaction in primary healthcare facilities.

Patients who attended the hospital for dental-related conditions were significantly more likely to report higher satisfaction with care compared with those presenting with digestive system problems. One possible explanation is that dental treatments often provide immediate relief from acute pain or visible improvement, which may produce more noticeable and emotionally rewarding outcomes than treatments for chronic or less symptomatic conditions such as digestive disorders. Supporting this interpretation, a study reported that 87% of dental patients in an emergency department experienced complete pain relief within 24 hours after treatment [32]. Another factor that may contribute to the higher satisfaction observed among dental patients is the cost and accessibility of dental services at the Savar Upazila Health Complex. Dental care at government facilities is highly subsidized; according to a national media report, outpatient dental services are provided for approximately BDT 10 per patient (≈ USD 0.04), with treatment materials supplied by the hospital [33]. In contrast, patients seeking treatment for other medical conditions may incur additional out-of-pocket expenses because certain medications are not always available at the facility, which may negatively influence satisfaction. Previous studies have also shown that lower treatment costs are associated with higher treatment satisfaction [34], while high treatment costs can lead to treatment avoidance, as observed in a Canadian population-based survey of dental care utilization [35]. Therefore, differences in both financial burden and immediacy of treatment outcomes may partly explain the higher satisfaction levels reported by dental patients in the present study.

This study found that the availability of comfortable seating in the waiting area was associated with more than a twofold increase in the likelihood of patient satisfaction. This finding highlights the important role of the physical environment of healthcare facilities in shaping patient experiences. Patients’ perception of waiting is influenced not only by the duration of waiting time but also by how they are treated during that period, including respectful triage, clear communication, and maintenance of privacy [36]. Previous research indicates that poorly designed waiting rooms, particularly in resource-constrained settings, may exacerbate patient stress through environmental stressors such as overcrowding, inadequate lighting, excessive noise, and confusing spatial layouts [37]. In contrast, thoughtfully designed environments that incorporate natural light, calming aesthetics, ergonomic furnishings, and biophilic elements can reduce stress and contribute to a more positive patient experience [37]. A study conducted in China examining the physical environment of outpatient waiting areas found that features such as adequate space, clear signage, pleasant visual environments, sound control, views of nature, and comfortable seating were strongly associated with higher patient satisfaction, shorter perceived waiting times, and more positive patient behaviors [38]. Moreover, satisfaction with waiting environments may vary across patient groups; factors such as age, gender, time of visit, and familiarity with the healthcare facility can influence how waiting spaces are perceived. For example, women and first-time visitors may report lower satisfaction due to discomfort or uncertainty within unfamiliar environments [39]. These findings highlight the importance of well-designed, patient-centered waiting areas as a practical and cost-effective strategy for improving patient satisfaction and perceived quality of care.

Access to healthy food within the healthcare facility was identified as a significant predictor of patient satisfaction in the studied Upazila Health Complex, with patients reporting such access having approximately 2.5 times higher odds of satisfaction. Adequate and nutritious food plays an important role in patient well-being, particularly for inpatients, as it supports recovery and contributes to the effectiveness of medical treatment [40, 41]. For outpatients as well, especially in facilities such as the Savar Upazila Health Complex where visits may last several hours for registration, consultation, and medication collection, access to safe and nutritious meals or snacks can improve comfort during the care process. Previous studies also support the relationship between food quality and patient satisfaction. For instance, an observational study reported that satisfaction improved when hospital food met patients’ expectations [42], while another study found that many inpatients experienced hunger due to factors such as inappropriate meal timing, physical access difficulties, limited food choices, and dissatisfaction with food quality [43]. Similar findings were reported in a hospital-based study in Greece, where food quality was identified as a key contributor to overall patient satisfaction [44]. Within the Bangladeshi context, earlier assessments have highlighted persistent shortcomings in hospital food services, including nutritionally inadequate meals, untrained food service staff, limited therapeutic diets, and poor kitchen sanitation [45]. More recent media investigations have echoed these concerns, reporting poor food quality, lack of mechanisms to tailor meals according to patients’ nutritional needs, and limited involvement of dietitians in hospital food management [46, 47]. These sources are included to provide contextual insights into hospital food practices in Bangladesh, where peer-reviewed evidence remains limited. Studies have also reported that many patients lack the option to choose their meals and frequently experience hunger between scheduled meal times [48].

Access to free medication emerged as an important factor associated with higher patient satisfaction in this study. This relationship may partly reflect the socioeconomic profile of the participants, as more than half of the respondents belonged to low-income households and may therefore face considerable financial barriers when purchasing prescribed medications. In the context of Bangladesh, this issue is particularly relevant because medicine availability in public healthcare facilities remains limited and stock-outs of essential drugs are frequently reported [49]. As a result, patients are often required to purchase medications out-of-pocket, contributing to the country’s high private health expenditure—estimated at approximately 67% of total health spending—which further exacerbates inequities in healthcare access [49]. Evidence from other low- and middle-income countries also highlights the importance of medicine affordability in shaping patient experiences of care. For example, a study conducted at Deder General Hospital in eastern Ethiopia reported that patients who received their prescribed medications from the healthcare facility were significantly more likely to report satisfaction with services, particularly among uninsured patients [50]. Similarly, research examining the national free pharmaceutical program in Sierra Leone reported high levels of patient satisfaction, underscoring the role of medicine affordability in improving perceptions of healthcare services [51]. Within the Bangladeshi context, previous studies have also indicated that the unavailability of free medicines and the burden of high drug costs remain important sources of dissatisfaction among patients attending public healthcare facilities [25].

The behavior of nurses emerged as an important determinant of patient satisfaction in this study. Effective nurse–patient communication—including attentive listening, clear explanations, and compassionate bedside interactions—has consistently been associated with higher patient satisfaction, better treatment adherence, and improved safety culture in healthcare settings [10, 52–54]. In resource-limited contexts such as Bangladesh, where infrastructural and resource constraints may affect service delivery, the interpersonal and communication skills of nursing staff become particularly important in shaping patients’ experiences of care [55]. Previous studies in Bangladesh have reported similar findings, indicating that both the quality of services at registration points and the behavior of nurses are significant predictors of patient satisfaction in government-financed healthcare programs [16]. Other evidence also highlights the importance of respectful and professional interaction between healthcare staff and patients; for example, one study reported that 63% of patients expressed satisfaction with hospital staff courteousness [56]. Improvements in nursing behavior have also been linked to broader health system reforms. In Iran, the implementation of the Health System Transformation Plan was associated with notable improvements in satisfaction with nursing care, accompanied by an increase in overall patient satisfaction scores [57]. Similarly, a large cross-sectional survey conducted across the United States and several European countries found that patient satisfaction was strongly influenced by nurses’ work environments, patient-to-nurse ratios, and the alignment between nurse-reported care quality and patients’ perceptions of hospital services [7].

Several variables did not remain statistically significant in the adjusted model, including prescription privacy, admission status, current residence, and distance from the hospital. The lack of association with prescription privacy may reflect the context of public healthcare facilities in Bangladesh, where consultations often occur in busy environments with shared spaces and rapid service delivery. Similarly, admission status did not appear to influence satisfaction, possibly because inpatient and outpatient services at the Upazila level share similar infrastructure, staffing, and care processes. Neither current residence nor distance from the hospital independently predicted satisfaction, suggesting that transportation convenience, rather than geographic proximity alone, may play a more important role in shaping patient experiences once access to care is achieved. Although waiting time was significant in the bivariate analysis, it did not remain significant after adjustment, indicating that other factors—such as the physical environment of the facility and staff behavior—may have a stronger influence on overall patient satisfaction.

From a machine-learning perspective, the findings indicate that tree-based ensemble models, particularly Random Forest and XGBoost, demonstrated strong predictive capability for modeling patient satisfaction in survey-based healthcare data. In the present analysis, Random Forest achieved the highest discrimination (ROC-AUC ≈ 0.82) with the lowest log-loss, while XGBoost produced the best overall classification balance (F1-score ≈ 0.87) with high recall (> 0.92), indicating strong ability to correctly identify satisfied patients. These results suggest that ensemble methods are well suited to capturing complex and potentially non-linear relationships among patient experience variables. Feature-importance analysis further highlighted that interpersonal and facility-related factors—including receptionist and nurse behavior, waiting-room comfort, sanitation, availability of food and medicines, and transportation convenience—were among the most influential predictors of patient satisfaction. Similar patterns have been observed in previous machine-learning studies of patient-experience data. For example, analyses using Hospital Consumer Assessment of Healthcare Providers and Systems datasets reported Random Forest models achieving AUC values above 0.80 when predicting patient satisfaction related to communication and care experience [58]. Other studies have also shown that Random Forest models can outperform traditional approaches such as logistic regression and naïve Bayes in satisfaction classification tasks, particularly when predictor relationships are non-linear or interactive [59]. In addition to predictive performance, tree-based models can reveal actionable determinants of patient experience; prior survey-based analyses similarly identified provider behavior and service processes as dominant predictors of satisfaction outcomes [60]. These findings indicate that machine-learning approaches, particularly ensemble tree models, can complement conventional statistical analyses by improving predictive performance while providing interpretable insights into the complex determinants of patient satisfaction.

Based on the study findings, several targeted interventions may help improve patient satisfaction in public healthcare facilities such as the Savar Upazila Health Complex. Improving physical accessibility to healthcare services remains essential, particularly through strengthening transportation connectivity and exploring decentralized service delivery models to reduce travel burdens for patients. The hospital environment should also be enhanced by ensuring clean, well-ventilated, and comfortable waiting areas with adequate seating arrangements, as physical comfort during waiting periods was identified as an important determinant of satisfaction. Ensuring reliable availability of free and essential medicines through improved procurement and supply chain management is equally important, particularly for patients from low-income households who may otherwise face financial barriers to treatment. The availability of safe and nutritious food within hospital premises should also be improved, especially for patients and attendants who spend extended periods in the facility. Furthermore, strengthening interpersonal aspects of care is critical; continuous training for healthcare staff—particularly nurses—in effective communication, empathy, and patient-centered care practices may help foster more respectful and supportive interactions with patients. Finally, institutionalizing routine patient satisfaction assessments could enable healthcare managers to systematically monitor patient experiences and identify areas for service improvement. Together, these measures may contribute to improving both patient satisfaction and the overall quality of care in primary healthcare facilities in Bangladesh.

This study has several strengths. It provides one of the few empirical assessments of patient satisfaction in a primary healthcare facility in Bangladesh using the PSQ-18 instrument, allowing comparison with international studies. The study also incorporated a relatively large sample of service users and included both outpatient and inpatient populations, offering a comprehensive picture of patient experiences within the facility. In addition, the integration of machine-learning techniques alongside conventional statistical analyses allowed the identification of complex relationships among satisfaction determinants and provided complementary insights into the relative importance of different factors influencing patient satisfaction.

Despite its contributions, this study has several limitations that should be considered when interpreting the findings. First, the cross-sectional design limits the ability to establish causal relationships between the identified determinants and patient satisfaction. Second, the study was conducted at a single primary-level public healthcare facility using a convenience sampling strategy, which may limit the representativeness of the sample and the generalizability of the findings to other healthcare settings in Bangladesh. Third, patient satisfaction and facility-related variables were measured using self-reported responses, which may be subject to recall bias and social desirability bias. Fourth, the inclusion of participants under 18 years of age and their analysis within the same category as older participants may introduce potential bias. Younger participants may be more susceptible to acquiescence bias, which may have influenced the validity of satisfaction-related responses despite efforts to ensure independent answering. This may have led to an overestimation of satisfaction levels among younger participants and should be considered when interpreting age-related findings. Future studies should consider age-stratified analyses to better account for potential differences in response patterns between pediatric and adult participants. In addition, the sample included a higher proportion of female participants, reflecting healthcare utilization patterns at the facility, which may have influenced overall satisfaction estimates. Finally, although the PSQ-18 questionnaire was translated into Bangla using a forward–backward translation procedure, a formal psychometric validation of the translated instrument was not conducted. Future studies using multicenter designs, probabilistic sampling methods, and longitudinal or mixed-method approaches would help validate and extend these findings across broader healthcare contexts in Bangladesh.

Conclusions

This study provides new evidence on the determinants of patient satisfaction at a primary healthcare facility in Bangladesh using both conventional regression and machine-learning approaches. The findings demonstrate that satisfaction is shaped by structural and interpersonal dimensions of care. Key predictors included convenient transportation, receipt of free medicines, comfortable waiting-room conditions, and nursing staff behavior; with these findings further corroborated by Random Forest feature-importance analysis. Within the studied upazila-level setting, these results suggest that improving patient satisfaction requires targeted improvements in physical accessibility, facility environment, availability of essential medicines, and patient-centered communication.

To translate these findings into practice, healthcare managers and policymakers should consider strengthening transportation connectivity to public facilities, upgrading waiting-area infrastructure, ensuring consistent free medicine supply through improved procurement, and institutionalizing structured communication training for nursing and reception staff. Routine patient satisfaction monitoring using standardized tools would further support evidence-based quality improvement. Importantly, the findings of the study should be interpreted with caution. The study was conducted at a single facility using convenience sampling, limiting generalizability beyond similar primary healthcare contexts in Bangladesh. The cross-sectional design precludes causal inference, and social desirability bias cannot be excluded. Multicenter, longitudinal studies using probability-based sampling are recommended to validate and extend these findings.

Author contributions

AAH and MTI conceptualized and designed the study. AAH led the overall implementation of the research project, including the development of data collection instruments, coordination of field activities, data collection, and data entry, with support from MTI, MAM, and FAM. AAH conducted the statistical analyses using SPSS and performed the machine-learning analyses with MEH. The analytical outputs and interpretation of the results were reviewed and validated by MEH, FAM, MAM, and MTI to ensure methodological rigor and accuracy. MTI provided primary supervision throughout the study, including guidance on study design, analytical approach, and interpretation of the findings, while FAM, MMA, and MAM contributed additional supervision and methodological support. AAH prepared the initial draft of the manuscript. MEH, FAM, MMA, MAM, and MTI contributed to the revision of the manuscript by providing substantial revisions, critical feedback, and intellectual input. AAH revised the manuscript with contributions from MEH, MMA, and MAM under the supervision of MTI. All authors reviewed and approved the final version of the manuscript.

Funding

Dr ALmerab is currently receiving funding support from the Princess Nourah Bint Abdulrahman University Researchers Supporting Project (Number PNURSP2026R563), Princess Nourah Bint Abdulrahman University, Riyadh, Saudi Arabia.

Data availability

The datasets will be made available to appropriate academic parties upon request from the corresponding author.

Declarations

Ethics approval and consent to participate

The study was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki (2013). Ethical approval was granted by the Thesis Proposal Approval Committee (2024), Department of Government and Politics, Jahangirnagar University, following evaluation of the research proposal on 4 December 2024. Written informed consent was obtained from all participants prior to data collection. For participants aged 10 to below 18 years, parental or guardian consent was obtained in accordance with ethical requirements. Permission to conduct the study was also obtained from the hospital authorities. Participants were informed about the study’s purpose, procedures, and their right to withdraw at any time without consequences. Participation in the study was voluntary, and no incentives were provided. To ensure confidentiality, all collected data were anonymized and securely stored throughout the study, and the data were used solely for research purposes.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.