The interplay between sleep quality, daytime fatigue, and daytime sleepiness is critical for understanding and diagnosing sleep disorders. Sleep quality—a multifaceted concept encompassing depth, continuity, and restorative nature of sleep [1]—is commonly assessed through polysomnography (PSG) or questionnaires [2]. Daytime fatigue significantly impacts quality of life [3], while daytime sleepiness reflects the propensity for unintended sleep episodes.

Past studies have shown inconsistencies between subjective experiences and objective measures. For example, Zavecz et al [4] linked self-reported sleep quality to cognitive performance, while O’Donnell et al [5] reported poor alignment between PSG outcomes and patient-reported sleep quality. Similarly, Aurora et al [6] analyzed how well Epworth Sleepiness Scale (ESS) scores correspond to Multiple Sleep Latency Test (MSLT) values. In psychiatric populations, similar dissociations have been reported between perceived and physiologically measured fatigue, as demonstrated by Stanyte et al [7] who found that individuals with anxiety and mood disorders often exhibit mismatched subjective fatigue ratings and objective sleep parameters.

Critically, objective metrics such as sleep onset latency (SOL) and total sleep time (TST) require disorder-contextual interpretation. Valko et al [8] demonstrate that shortened SOL indicates pathological sleep-wake fragmentation in narcolepsy but signifies health in normal sleepers, while subjectively prolonged SOL in insomnia reflects hyperarousal despite normal PSG. Similarly, fatigue—though quantifiable via scales such as the Fatigue Severity Scale (FSS)—lacks consistent PSG correlates across disorders, complicating objectification[9].

Despite these assessments being routinely documented clinically, no study has leveraged real-world clinical records to analyze subjective-objective alignment across disorders.

To address this gap, we propose a novel natural language processing (NLP) framework for modeling subjective-objective alignment through aspect-based sentiment analysis. Our system extracts (1) patient-reported experiences, (2) objective test results (eg, MSLT and PSG), and (3) clinician commentary on alignment or mismatch between the two. This triadic analysis allows for identifying patterns of misperception—for instance, patients who perceive severe sleep onset problems despite normal PSG values.

Unlike prior work that focused on isolated biomarkers such as Apnea-Hypopnea Index (AHI) or ESS scores, our approach uses language models to extract detailed clinical reasoning from documentation. This allows identification of emerging “sleep-wake discrepancy” phenotypes, such as patients with objectively normal sleep but persistently poor subjective reports.

Sentiment analysis in medical and clinical text has received increasing attention over the past decade [10,11]. Clinical documents often contain detailed descriptions of a patient’s health status, including observations, diagnostic findings, and treatment plans. Analyzing this information is important for determining whether patient outcomes are improving or deteriorating, and for assessing the overall impact of a condition on the patient’s well-being. In addition to objective clinical data, patients’ self-reported experiences may also offer useful indicators that complement medical assessments.

Recent developments in pretrained language models, such as those introduced by Devlin et al [12], have enabled the use of BERT-based architectures for sentiment analysis across multiple languages and granularities. In our work, we define sentiment across 5 discrete levels, which facilitates scaling into a normalized range between 0 and 1. Behera et al [13] presented a model combining Word2Vec embeddings with convolutional and recurrent layers (convolutional neural network and long short-term memory), achieving 91.2% accuracy. Their method also includes polarity normalization to improve classification performance. Other established techniques for sentiment and aspect identification in English-language texts include BiLSTM [12,14], TextCNN [15], TextRNN [16], TextRCNN [17,18], and DPCNN [19].

With the emergence of transformer-based models, research has increasingly focused on controlling language model behavior during fine-tuning, inference, and output postprocessing [20]. For example, Ziegler et al [21] applied reinforcement learning (RL) with human feedback to improve output coherence, while Schulman et al [22] proposed the Proximal Policy Optimization (PPO) algorithm for stable policy updates. Pascual et al [23] introduced Keyword2Text, a framework that guides text generation based on semantic similarity and topic constraints. These control strategies can enhance performance in tasks such as sequence labeling and sentiment classification.

Despite these advances, the use of large language models (LLMs) for aspect recognition in German clinical text remains limited. In addition, data privacy regulations restrict the use of commercial cloud-based models in health care environments. To address these challenges, our study investigates the deployment of open-source language models in secure, on-premises settings. We evaluate their performance in aspect recognition using manually annotated German clinical data. Subsequently, we apply sentiment analysis using a fine-tuned BERT model and compare its performance to a German GPT-2 model, which we adapt through supervised learning and reinforcement-based posttraining techniques.

As can be seen in Table 1, prior work has explored sentiment in clinical text [10,11], or compared ESS to MSLT [6], but none have jointly modeled aspect-level sentiment, LLM-based extraction, and alignment between subjective and objective sleep data. Our contributions include:

The developed pipeline for aspect-oriented sentiment analysis is shown in Figure 1. The single steps are described together with the dataset in the following. Our approach involves extracting subjective perceptions of sleep quality, daytime sleepiness, and daily fatigue from the clinical documentation. Specifically, we consider the documented patient history descriptions and the reports related to the PSG examination.

Concurrently, objective benchmarks will be gathered from numerical results in objective sources, including MSLT, MWT (Maintenance of Wakefulness Test), P-AW (Actigraphy Wrist), and parts of the PSG report.

In a first step, entities and phrases referring to our 3 aspects of interest (day sleepiness, sleep quality, and fatigue) are extracted. Once these aspects are identified, we classify them according to the sentiment expressed for each aspect separately. The next step involves calibrating the objective numerical assessments to match the subjective sentiment scores, allowing for a direct comparison between subjective and objective data. Finally, by evaluating these 2 sets of sentiment values, we statistically analyze the extent of discrepancies in the diagnosis of sleep disorders, providing insights into how patient perceptions align with clinical assessments.

Our dataset originates from inpatient records created at the sleep laboratory of the University Hospital Bern, covering treatments between 2000 and 2021. This collection is part of the Bern Sleep Registry [25]. Its secondary use received approval from the Cantonal Ethics Committee (KEK-Nr. 2022‐00415: “Bern Sleep Registry: the sleep disorder patient cohort of the Inselspital, University Hospital Bern”).

The database comprises German-language coded clinical records of patients diagnosed with sleep disorders. Over 10,000 of these records have been retrospectively and meticulously categorized by physicians specialized in sleep disorders based on the International Classification of Sleep Disorders, Third Edition (ICSD-III). Table 2 shows the types of text documents that are considered from this database for this paper.

For various validations, we have prepared the following manually annotated datasets to validate and fine-tune the pretrained model:

The 2000 sentences were sampled from clinical documents belonging to 100 unique patients, spanning multiple sleep disorder categories. These sentences were drawn proportionally from patient histories, diagnostic summaries, and follow-up reports to ensure a representative mix across conditions and report types. Stratified sampling was used to balance diagnoses, age groups, and sentiment distribution. To avoid patient-level bias, no more than 10 sentences per patient were included in the annotated set.

To ensure robust analysis of subjective-objective alignment in daytime sleepiness, we included only those patients who had complete paired data for the following instruments: Karolinska Sleepiness Scale (KSS)—subjective measure, MSLT—objective physiological measure, and MWT—behavioral sleepiness measure. This resulted in a final cohort of 100 patients with complete records across all 3 modalities, allowing for valid computation of discrepancy and misperception scores. For other dimensions, such as fatigue and sleep quality, which currently lack standardized objective clinical benchmarks, we retained a broader set of records (eg, 3815 entries for sleep quality) to support exploratory sentiment-based analysis.

To analyze sentiment by aspect, we annotated 50 representative sentences per aspect totaling 150 sentences. Table 3 summarizes the frequency of extracted entity groups by category, while Table 4 presents the sentiment annotations and related statistics.

In the following, we define the dataset and tasks considered in this paper.

Dataset Definition

Let the dataset be defined as:

D={(x_i, A_i, y_i)|=1,...,N

Where:

Aspect Extraction via LLMs

For each input sentence x_i and aspect category, for example, c ∈ C_Categories, we define a prompt P_c and extract:

a_i^{(c)=LLM (P_c,x_i)} for each aspect category c

The union of these yields the aspect set:

A_i=U_ca_i^(c)

To align with canonical medical terms, compute cosine similarity:

sim(e, c) = cosine(embedding(e), embedding(c))

Accept the match if sim≥threshold

Sentiment Classification (BiLSTM+Self-Attention)

Training Objective (Supervised Fine-Tuning: L1)

Use categorical cross-entropy for supervision:

Sentiment Control via GPT-2 (L2: Reinforcement Learning)[29]

After initial fine-tuning, a GPT-2 model is trained to modify or generate text while controlling for sentiment class y_i

Logit Biasing /Reward Shaping: to

Where r(xi^,yi) is a reward signal based on:

The full objective (SFT+RL) becomes:

where λ∈[0,1] is a hyperparameter controlling the influence of reinforcement-based alignment.

Normalization of Subjective and Objective Measures

Min-max normalization for raw clinical scores r over scale [S_min, S_max]:

Examples:

For aspect extraction, we define 4 distinct groups of clinical entities based on their frequency and contextual relevance, with the aim of analyzing discrepancies between subjective reports and objective clinical findings (see Table 2). While aggregating sentiment scores can offer a general view of polarity across different textual spans (eg, sentences or paragraphs), they are insufficient for capturing detailed sentiment variations related to specific medical aspects such as symptoms, complaints, diagnoses, or co-occurring conditions. To address this limitation, we implemented a fine-grained entity extraction approach that combines open-source LLMs with standard clinical entity recognition techniques.

Specifically, we use 2 self-hosted models: Falcon 40B [27] and Mixtral 8x7B [30]. Falcon 40B is a dense, decoder-only language model designed for a broad range of NLP tasks. Mixtral 8x7B, by contrast, follows a sparse mixture-of-experts architecture, activating 2 expert modules per token, thereby achieving a favorable balance between performance and computational efficiency. Despite having 46.7 billion parameters in total, only a subset is used at each step, making it suitable for resource-constrained environments. We evaluated Falcon 40B as a baseline but used Mixtral 8x7B for all final extractions reported here.

Aspect-level sentiment scoring is achieved by linking the extracted entities to their corresponding sentiment values within a given sentence. This allows us to compute sentiment scores that are specific to individual clinical aspects. The entity extraction process is carried out in 2 stages: (Stage 1) Aspect identification using prompt-based querying with the locally hosted Mixtral 8x7B model. Separate prompts are constructed for each category—symptoms, complaints, diagnoses, and patient feedback—to ensure focused and interpretable extraction.

To extract patient feedback entities, we design targeted natural language prompts structured to guide the model toward recognizing relevant subjective expressions. Below is a representative prompt used with the Falcon 40B model:

To enable comparability across heterogeneous clinical measurements and support sentiment interpretation, we apply min-max normalization to both objective and subjective metrics. This ensures all scores lie within the interval [0, 1], where 0 indicates maximal symptom burden and 1 indicates no impairment. This transformation provides a unified scale for downstream sentiment mapping.

For objective physiological measures such as PSG-recorded sleep latency, we applied the following normalization: Normalized Latency Score=1–(Raw Value/Max Latency) where the clinically derived maximum latency was set to 120 minutes, in accordance with standard clinical PSG protocols. This formulation ensures shorter latency (indicative of better sleep initiation) corresponds to higher normalized scores.

We deploy a BERT-based BiLSTM architecture to perform sentence-level sentiment classification across 5 polarity levels (see Figure 2). The model ingests token sequences up to 250 tokens, with BERT embeddings (768 dimensions) [31] providing context-aware representations.

We fine-tune both multilingual BERT and XLM-RoBERTa [32] on a corpus of 1200 German clinical sentences annotated with 5-stage sentiment labels, derived from Mixtral 8X7B prompting and manual correction (interannotator agreement >89%).

Negated expressions in German are normalized into unified tokens prior to BERT encoding to stabilize polarity signals. The resulting embeddings are processed by a BiLSTM layer followed by sequential feature aggregation. The output logits are mapped to 5 sentiment classes, which are rescaled to a [0, 1] range:

0 (very negative), 1 (negative), 2 (neutral), 3 (positive), 4 (very positive)

The model is trained using a 60/20/20 split across training (1200), validation (400), and test (400) sets.

Ethical approval for this study was obtained from the Kantonale Ethik Kommission Bern (Cantonal Ethics Committee Bern) for multiple project components: Project part 1: SNS Project (2000-2016), BASEC-ID 2016-00409 and Project part 2: Bern Sleep Registry (“The sleep disorder patient cohort of the Inselspital, University Hospital Bern”), KEK-Nr. 2022-00415.

The secondary use of data from the Bern Sleep Registry was also approved by the Cantonal Ethics Committee. All data were handled in accordance with institutional and Swiss data protection regulations. Informed consent was obtained as required for each project component, and all participants were informed of their ability to opt out of data use. All data were fully de-identified prior to analysis to protect participant privacy and ensure compliance with applicable data protection standards.

This section presents the results of our aspect-based sentiment analysis pipeline, including [1] 5-stage sentiment classification using transformer-based models, and [2] LLM-based extraction of clinical entity aspects. The analyses were applied to a dataset of sleep-related medical texts and further supported by clinical insights into subjective-objective misperception.

Table 9 summarizes the performance of several models for 5-stage sentiment classification, evaluated using standard multiclass metrics. Initial models such as BERT and RoBERTa achieved moderate performance (accuracy 61% and 69%, respectively). Incorporating a BiLSTM layer improved performance to 78%. Further gains were achieved with a fine-tuned GPT-2 model. SFT alone yielded 81% accuracy, and RL with logit-space modulation increased it to 87%. These results demonstrate the effectiveness of multistage alignment in modeling sentiment nuances. Metrics reported in Table 9 are macro-averaged across 5 sentiment classes, based on an 80/20 stratified test split.

All metrics were computed using macro-averaging, giving equal weight to all sentiment classes. The AUC-ROC (area under the receiver operating characteristic curve) values were obtained using a one-vs-rest strategy, macro-averaged across 5 classes. These approaches ensure that both frequent and rare sentiment classes are equally represented in performance assessment.

Figure 4 visualizes the ROC curves of the models. GPT-2 variants with SFT+RL training consistently outperformed the baselines, reflecting enhanced class separability through reinforcement learning and logit regularization.

The ROC-AUC curves illustrate the one-vs-rest performance of 6 models on the 5-stage sentiment classification task using clinical text from sleep disorder records. All models perform substantially above the chance level (AUC=0.5), with performance steadily improving from BERT (AUC=0.76) to RoBERTa (AUC=0.84), and further to RoBERTa+BiLSTM (AUC=0.90). Fine-tuned GPT-2 (SFT) achieves strong performance (AUC=0.93), which is further enhanced by reinforcement learning (RL1 and RL2) and logit modification, reaching up to AUC=0.96. The stair-step shapes reflect the small test set (400 samples), but the consistent trend shows that reinforcement learning significantly boosts the model’s ability to distinguish fine-grained sentiment levels, supporting its potential for clinical decision support in contexts such as misperception of sleepiness.

Clinical entity extraction was performed using prompting strategies applied to Falcon 40B and Mixtral 8X7B models. Table 10 compares their performance on 4 annotated categories: Symptoms, Diseases, Complaints, and Feedback. Mixtral 8X7B showed strong performance with macro F₁ of 0.8490, outperforming Falcon 40B (macro F₁ of 0.7265).

Since Mixtral 8X7B consistently outperformed Falcon 40B across all evaluated categories, we applied Mixtral 8X7B for all entity extraction tasks in our sentiment analysis pipeline. Beyond the benchmark evaluation, the model extracted a total of 873 unique symptom terms, along with 59 distinct diseases, 432 complaints, and 224 feedback-related entities across the full clinical corpus.

This study introduced a sentiment analysis framework for clinical sleep narratives, revealing key insights into both method performance and clinical relevance. The BERT-BiLSTM architecture improved sentiment accuracy by addressing negations and drift, while the Mixtral 8X7B model outperformed Falcon 40B in aspect extraction due to better handling of complex German syntax. A fine-tuned GPT-2 model with reinforcement learning achieved high sentiment classification accuracy (87%) and offered a resource-efficient alternative to LLMs. Clinically, sentiment-derived scores revealed consistent misperception patterns—such as underestimated sleep latency and overestimated duration—highlighting the importance of aligning subjective and objective measures. These findings support the integration of sentiment-informed misperception analysis into sleep medicine workflows to improve diagnosis, treatment selection, and patient safety.

In this study, we introduced and evaluated an aspect-based sentiment analysis approach tailored to clinical narratives in sleep medicine. Several key observations emerged from our findings. First, when mapping sentiment model outputs to scalar indicators of sleep disorders, it became necessary to adjust for negations and drifted sentiment—especially in baseline BERT models. Phrases that negated the presence of symptoms or diagnoses were frequently misinterpreted as negative sentiment. This issue was mitigated using a BERT-BiLSTM architecture, which improved performance through representation adaptation and fine-tuning.

For aspect (entity) extraction using prompted LLMs, recall was generally low, particularly for feedback-related content. In contrast, precision was high for identifying symptoms, diseases, and complaints. The poor recall is primarily attributed to the token limit and the generation style of Falcon 40B under one-shot prompting. Interestingly, few-shot prompting increased precision but further reduced recall and F₁-score. This reveals a trade-off that may be addressed either by increasing computational capacity (current setup: 128 GB Nvidia DGX server) or optimizing prompting strategies.

The Mixtral 8X7B model significantly outperformed Falcon 40B, delivering higher precision and recall for aspect recognition, even under single-shot settings. Its ability to handle longer input contexts and complex linguistic structures—such as German postnegation—proved particularly advantageous in the clinical domain. These results support the use of Mixtral 8x7B for German-language medical NLP tasks where nuanced comprehension is critical.

Compared to BERT-BiLSTM, a fine-tuned GPT-2 model trained on 1200 sleep-related clinical sentences achieved comparable performance (accuracy: 81%). Incorporating keyword-guided SFT and 2 RL epochs with logit modification raised accuracy to 87%. This dual-phase tuning strategy enhanced the model’s flexibility and control. It also outperformed sequential prompting approaches commonly used with larger LLMs—offering higher precision with lower resource requirements, particularly important in privacy-sensitive clinical settings.

The interpretation of sentiment-derived misperception patterns must consider both the nature of subjective reporting and the inherent limitations of objective measurement in clinical sleep data. Misperception—defined as a mismatch between how patients feel and how their physiological state is measured—varies across dimensions of sleep disorders, with important diagnostic and therapeutic implications.

A key finding in our study was the divergence in sentiment-derived scores for similar clinical constructs. For instance, patients frequently underestimated their sleep latency (eg, reporting quick sleep onset) yet simultaneously overestimated their sleep duration, contrary to PSG and MSLT results. This subjective-objective mismatch—commonly referred to as sleep misperception—was particularly evident across latency, duration, and fatigue indicators. While traditional statistical tests may not fully capture the nuances of sentiment-informed scores, the consistent directionality and magnitude of these mismatches highlight their clinical importance.

However, the objectification of sleep quality and fatigue is inherently complex. Short sleep latency, often seen as a marker of healthy sleep, can paradoxically indicate pathological hypersomnolence in conditions such as narcolepsy. Similarly, fatigue—unlike sleepiness—lacks a gold-standard physiological test and is primarily self-reported, making its objectification more elusive. These nuances emphasize the need to interpret subjective scores considering disorder-specific clinical contexts.

In contrast, daytime sleepiness offers a more quantifiable target for misperception analysis. Here, subjective ratings such as the KSS and text-derived sentiment scores can be directly compared with objective markers such as MSLT latency. The KSS is particularly useful in this context, as it captures state-based, momentary sleepiness, allowing for meaningful trial-by-trial comparisons. For example, KSS values close to 0 (indicating “extremely sleepy”) should theoretically align with MSLT latencies under 8 minutes. In our analysis framework, misperception magnitude increases when a patient reports low KSS (ie, “did not fall asleep”) while objective data confirm they did—for example, falling asleep within 10‐20 minutes in MSLT trials. This alignment enables fine-grained modeling of momentary misperception and supports targeted subanalysis of repeated MSLT and KSS pairs (eg, mslt_kssdur_10 vs mslt_sleeplatency_10).

From a clinical perspective, understanding the type and direction of misperception has actionable relevance.

For instance, in insomnia, patients often report exaggerated sleep difficulties (eg, longer sleep latency than measured). These cases may benefit more from cognitive behavioral therapy for insomnia rather than pharmacotherapy. Discussing the misperception with the patient is often therapeutic. However, because objective tests such as PSG are rarely used in routine insomnia management, wearable technologies (eg, smartwatches) could serve as surrogate tools to estimate sleep latency, duration, and fragmentation—offering scalable insight into real-world misperception.

In hypersomnia or sleep-disordered breathing, the opposite trend may occur—patients feel they slept well but exhibit abnormal sleep architecture or excessive daytime sleepiness in tests. In such cases, unrecognized sleepiness can elevate safety risks, such as motor vehicle or occupational accidents. Early identification of these cases using sentiment-objective comparison could trigger earlier clinical intervention or patient education.

Altogether, our sentiment analysis framework, enriched with structured objective metrics, offers a promising route for personalized assessment of sleep misperception. By comparing KSS, ESS, and text-derived sentiment scores against MSLT latency across multiple configurations, we can determine which input combination best captures the true subjective state. This multipronged analysis will inform future iterations of our clinical decision-support tools and help stratify patients for tailored treatment approaches.

This study demonstrates that sentiment analysis applied to structured clinical narratives can uncover meaningful patterns in how patients perceive and report sleep disorders—especially in identifying discrepancies between subjective self-reports and objective assessments. Among the models evaluated, the BERT-BiLSTM architecture showed strong performance in domain-specific sentiment detection. However, the fine-tuned German GPT-2 model, enhanced with supervised and reinforcement learning (PPO), achieved the highest accuracy for 5-stage sentiment classification, offering an effective balance between adaptability and computational efficiency.

Our findings highlight the clinical value of sentiment-based misperception modeling. For example, patients with insomnia often overestimate wakefulness, while those with hypersomnia or sleep apnea tend to underreport daytime sleepiness. The use of reinforcement learning techniques proved especially effective for capturing emotionally nuanced language in clinical narratives while maintaining sensitivity to domain-specific clinical distinctions.

However, this study is not without limitations. All data were drawn from a single clinical registry, which may limit the generalizability of findings across institutions or populations. Furthermore, the annotated training set for sentiment and entity extraction was relatively small, potentially constraining model robustness and the diversity of learned patterns.

In future work, we aim to expand the dataset by incorporating records from additional clinical centers and to enrich the annotation set to enable more nuanced model training and validation. Sentiment modeling will be extended to cover additional clinical entities such as pain and cognitive symptoms, while correlations between sentiment-derived misperception scores and treatment outcomes or resistance will be further explored. We also plan to integrate physiological features—such as microarousals and sleep stage transitions—to improve subtype classification and to leverage wearable sensor data to validate and calibrate sentiment-based misperception metrics. Ultimately, our goal is to develop a precision sleep medicine framework that integrates objective physiological markers with patients’ emotional and cognitive interpretations of their symptoms, captured through advanced sentiment and entity modeling of clinical narratives.