A Narrative-Driven Computational Framework for Clinician Burnout Surveillance

Alyssa Meczkowska; Fazel Keshtkar; Syed Ahmad Chan Bukhari

REVIEW 4 major objections 5 minor 19 references

A Narrative-Driven Computational Framework for Clinician Burnout Surveillance

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper argues that ICU discharge notes encode clinician burnout signals strong enough for a provider-level classifier to reach F1 0.84.

desk verdict The burnout label is built from the same narrative features the model predicts, so the F1=0.84 is circular; the paper is an honest, reproducible proof-of-concept but not a valid burnout detector. read the letter →

arxiv 2509.04497 v1 pith:ZMRBHLFX submitted 2025-09-01 cs.CL cs.AI

Syed Ahmad Chan Bukhari , Fazel Keshtkar , Alyssa Meczkowska This is my paper · ORCID

classification cs.CLcs.AI

keywords clinicianburnoutMIMIC-IVICUdischargesummariesBioBERTlatentDirichletallocationlogisticregressionclinicalNLPproviderwell-being

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that the free-text narratives clinicians write in ICU discharge summaries carry detectable, quantifiable signals of burnout that standard survey and metadata-only methods miss. It combines sentiment scores from a clinical BioBERT model, counts of stress-related phrases, topic weights from a five-topic LDA model, and workload proxies derived from EHR tables, then feeds these into a provider-level logistic regression. On a held-out 20% of 973 providers, the classifier gets precision 0.80, recall 0.89, and F1 0.84, beating metadata-only baselines by at least 0.17 F1. The authors claim this makes routine discharge notes a viable substrate for proactive clinician well-being monitoring, with Radiology, Psychiatry, and Neurology showing the highest risk.

What carries the argument

The load-bearing object is the silver-standard burnout label and the feature stack that predicts it. The label is a threshold rule applied to the narrative text: at least 12 high-confidence negative sentences from a fine-tuned BioBERT sentiment model and at least 7 hits in a handcrafted stress-cause lexicon. The classifier then learns to reproduce that label from a fused vector of BioBERT sentiment scores, lexicon counts, five-topic LDA weights, and workload proxies (lab orders, procedures, mortality flag, length of stay). Because the same narrative measurements define the target and supply the features, the pipeline's 'burnout' is an internally consistent text-based construct, not an extern

What would settle it

Administer a validated burnout survey (e.g., Maslach Burnout Inventory) to the same 973 providers, or collect discharge notes plus concurrent survey scores in a new cohort, and check whether the ≥12-negative-sentences / ≥7-cause-mentions rule assigns true burnout cases with better-than-chance accuracy. If the survey-positive group is not enriched among flagged providers (e.g., AUC around 0.5), the paper's central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that clinician burnout leaves a measurable trace in ICU discharge summaries, and that a provider-level classifier can read that trace. The authors define a 'silver-standard' burnout label from narrative output itself: a clinician is flagged when their notes contain at least 12 high-confidence negatively scored BioBERT sentences and at least 7 mentions of stress-related causes such as overtime or short-staffing. On a held-out 20% of 973 providers, logistic regression trained on fused sentiment, lexicon, topic, and workload features reaches precision 0.80, recall 0.89, F1 0.84, beating metadata-only baselines by 0.17 F1 or more. The same proxy places Radiology, Psy

Load-bearing premise

The claim stands or falls on whether writing at least 12 highly negative sentences and at least 7 stress-related phrases in discharge notes is actually a sign of burnout; the paper has not yet tested this rule against a validated burnout survey.

Editorial extensions

If this is right

If the narrative signal is real, hospitals could monitor burnout risk continuously by scoring discharge notes as they are written, rather than waiting for survey cycles.
The feature mapping to MBI dimensions gives administrators a reason language: high first-person pronouns and negative sentiment (emotional exhaustion), stress-lexicon hits (depersonalization), and administrative-topic weight (reduced accomplishment).
The specialty-level ordering suggests intervention resources could be targeted to Radiology, Psychiatry, and Neurology first.
The claimed ≥0.17 F1 gain over metadata-only models implies that note text adds predictive information not available in clickstreams or order counts.
The authors' stated next step, prospective comparison against survey scores, would turn the proxy into a screening instrument if it agrees.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the burnout label is derived from the same narrative features the classifier consumes, the reported F1 likely overstates real-world detection of clinical burnout; only an independent survey-based label can settle the gap.
The heavy concentration of notes among a few providers (top 10 author 130–170 notes each) means provider-level averages are shaped by a small number of high-volume authors; the specialty ranking could reflect documentation load rather than distress.
Operationalizing this as surveillance of clinicians' free text raises consent and trust issues the paper does not discuss; deployment would require transparent governance.
If the narrative markers are leading indicators, one testable extension is to track whether flagged providers go on to take leave, reduce hours, or score positive on a later survey.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Short version: the central result doesn't survive contact with its own methods. The "burned-out" label is defined in Section III.A as providers with at least 12 high-confidence BioBERT sentences and at least 7 cause-lexicon mentions; the classifier then uses the same sentiment scores and lexicon counts as predictors. So the F1 of 0.84 largely measures how well logistic regression recovers its own threshold rule. The paper even concedes in Section IV.E that the label has not been benchmarked against MBI or PFI. Without external validation, the specialty ranking and the surveillance claim rest on the same unvalidated proxy.

That said, there is something here worth acknowledging. The idea of using narrative features from ICU discharge notes for burnout surveillance is a reasonable application, and the authors are transparent about their pipeline. The MIMIC-IV data handling and feature engineering are described in enough detail to reproduce. The related work table is useful, and the mapping to MBI dimensions is thoughtful, even if speculative. The paper is clearly written and the limitations section is honest.

The soft spots are load-bearing. First, the circularity makes the F1 meaningless as evidence of burnout detection. Second, with 43 positive providers (~4.4%), a 20% test set leaves roughly 8–9 positives, so the performance estimate is very noisy. Third, the abstract claims the classifier beats metadata-only baselines by at least 0.17 F1, but those baselines are never shown. Fourth, the specialty analysis uses the same proxy, so the Radiology/Psychiatry/Neurology ranking inherits the same problem. The authors acknowledge the lack of psychometric validation but frame it as future work; fair, but the current conclusions should be correspondingly modest.

Who is this for? Someone working on clinical NLP or clinician well-being might find it a useful starting point, and it serves as a teaching example of label leakage in health text mining. It does not deserve publication as a claim of burnout detection. With real MBI or PFI labels, or at least independent annotation of a subsample, the circularity could be broken and the approach might have something.

Recommendation: I would not desk reject—the topic is important and the authors are clearly serious—but I would send it to reviewers with a strong directive that the current labeling is circular and the baselines are missing. In its present form the central claim is unsupported.

Referee Report

4 major / 5 minor

Summary. The paper proposes a computational pipeline to detect clinician burnout from ICU discharge summaries in MIMIC-IV. It combines BioBERT sentence sentiment scores, a handcrafted stress lexicon, LDA topic weights, and workload proxies, then trains a provider-level logistic regression classifier on these features. Providers are labeled as burned out if they produced at least 12 high-confidence BioBERT sentences and at least 7 cause-lexicon mentions. The authors report precision 0.80, recall 0.89, F1 0.84 on a 20% held-out set, claim this surpasses metadata-only baselines by at least 0.17 F1, and identify Radiology, Psychiatry, and Neurology as high-risk specialties. The central claim is that clinical narratives contain actionable signals for proactive burnout surveillance.

Significance. If the label were valid, the paper would offer a scalable, passive burnout surveillance method using routinely available EHR narratives, complementing or replacing survey-based tools. The use of a public dataset (MIMIC-IV), transparent feature groups, and a modest, interpretable classifier are positive features. However, the significance is conditional on the validity of the burnout label, and the manuscript provides no external criterion validation. Because the label is a hand-made threshold on the same narrative features later used as predictors, the headline results may only demonstrate that logistic regression can recover that threshold. The paper's contribution is therefore not established as stated.

major comments (4)

[Section III.A and III.B] The burnout label is defined in Section III.A as providers with 'at least 12 high-confidence BioBERT sentences and at least 7 cause mentions.' Section III.B trains logistic regression on sentiment scores, lexicon counts, LDA topic weights, and workload proxies. The outcome is a deterministic function of the very sentiment and lexicon features used as predictors. The reported F1=0.84 on the held-out set therefore likely reflects the classifier's success at reconstructing the labeling rule, not at detecting a clinically meaningful burnout construct. This is a load-bearing circularity. To support the claim, the authors must either validate the label against an independent instrument (e.g., MBI) or demonstrate that the classifier identifies burnout beyond what the threshold rule already encodes.
[Section IV.E and IV.A] The paper concedes in Section IV.E that the labeling rule 'has not yet been benchmarked against established psychometric instruments such as the Maslach Burnout Inventory.' Given that the label is the only ground truth for the classifier, this is not a minor limitation but a direct threat to the central claim. The label's assumption that narrative counts (e.g., negative sentences, cause mentions) equal burnout is untested. Without criterion validity, the specialty ranking (Radiology, Psychiatry, Neurology) and the proactive surveillance conclusion inherit the same unvalidated proxy. The authors should provide external validation or substantially reframe the claims as measuring narrative stress signals, not burnout.
[Abstract and Section IV.C] The abstract and Section IV.C claim that the narrative-based classifier 'surpasses metadata-only baselines by ≥ 0.17 F1 score,' but no metadata-only baseline results are reported anywhere in the paper. There is no table, figure, or numeric comparison showing the metadata-only F1, precision, or recall. This makes it impossible to assess whether narrative features add signal beyond workload proxies or whether the F1 gain is simply an artifact of using the same features to define and predict the label. The comparison must be shown explicitly.
[Section IV.A] With 43 positive providers out of 973 (~4.4%) and a 20% held-out set, the test set contains roughly 8–9 positive providers. A precision of 0.80 and recall of 0.89 are based on a very small number of positives, so the F1 estimate has high variance. No confidence intervals are provided. The paper should include interval estimates or, at minimum, a confusion matrix and a sensitivity analysis of the threshold choice.

minor comments (5)

[Abstract] The claimed baseline improvement '≥ 0.17 F1' is missing the comparison data; see major comment. Also, the abstract mentions 'five-topic LDA' and 'workload proxies' without defining them, which is acceptable for an abstract but could be clarified.
[Section III.A] There is a typo: 'Sentence scores with a value of ≥ were considered positive' — the threshold value is missing. Relatedly, the positive sentiment distribution is reported in Section IV.A as '1.2%,' with incomplete text ('74.6'). These should be corrected.
[References] Reference [11] is cited to 'TechTarget' rather than the original HiPAL paper, and reference [14] uses 'Anonymous' instead of an author list. Please verify and complete the citations.
[Section II.B] The phrase 'exhibited a significant misclassification rate of 50' is incomplete; the sentence breaks off. This should be fixed.
[General] Figure and table cross-references are inconsistent in places (e.g., 'Table I' is mentioned before its position, and Figure captions in Section IV are not all referenced in the text). A proofreading pass would improve readability.

Circularity Check

2 steps flagged · score 8.0 of 10

The burnout label is a threshold on the same BioBERT sentiment and cause-lexicon features later used as classifier inputs, so the reported F1 largely measures recovery of the labeling rule rather than detection of burnout.

self definitional [Section III.A (Data and Feature Engineering); Section IV.A (Dataset Overview); Section III.B (Model Training & Evaluation)]
"Providers were flagged as “burned-out” if they produced at least 12 high-confidence BioBERT sentences and at least 7 cause mentions. ... We identified high-risk providers by evaluating their output, which included ≥ 12 negative sentiment notes and ≥ 7 cause-lexicon mentions. ... A logistic regression classifier trained on a fused feature set (sentiment scores, topic weights, lexicon counts, and workload proxies) demonstrated robust burnout detection."

The binary label Y is defined by thresholds on counts of (a) high-confidence/negative BioBERT sentiment sentences and (b) cause-lexicon mentions. The classifier's feature set explicitly includes sentiment scores (BioBERT-derived) and lexicon counts (cause mentions), plus topic weights and workload proxies. Thus Y is a deterministic function of a subset of X; logistic regression is essentially being asked to approximate that threshold rule. A high held-out F1 therefore reflects successful recovery of the paper's own threshold construction, not detection of a clinically validated burnout construct. The paper's own limitation (Section IV.E) concedes the labeling rule 'has not yet been benchmarked against established psychometric instruments such as the Maslach Burnout Inventory,' so no extern
fitted input called prediction [Section IV.D (Specialty-Level Analysis and Feature Importance)]
"Feature inputs to the logistic regression model were selected based on clinical relevance, including the provider’s count of high-severity notes, the number of “Long Hours” cause-lexicon mentions (e.g., “overtime”), and first-person pronoun frequency to capture workload, operational strain, and narrative tone in predicting burnout risk."

The label is defined by high-severity narrative output (≥12 high-confidence/negative BioBERT sentences plus ≥7 cause mentions). The model is then given the provider's 'count of high-severity notes' as a predictor of a label that is itself defined by high-severity note counts. This is a fitted input called a prediction: the count of high-severity notes is essentially the label variable fed back as a feature, so the reported precision/recall/F1 of 0.80/0.89/0.84 is inflated by construction. The remaining features (e.g., cause-lexicon mentions) are also the very counts used to construct the label.

full rationale

The central derivation chain is circular in its core. Section III.A sets the burnout flag as: 'at least 12 high-confidence BioBERT sentences and at least 7 cause mentions.' Section IV.A restates this as '≥ 12 negative sentiment notes and ≥ 7 cause-lexicon mentions.' The classifier trained in Section III.B uses 'sentiment scores, topic weights, lexicon counts, and workload proxies' as features, i.e., the same BioBERT sentiment and cause-lexicon families that define the label. A logistic regression can approximate a threshold on those inputs, so F1=0.84 largely quantifies self-prediction. Section IV.D makes this even more direct by listing 'the provider’s count of high-severity notes' as a predictive feature, which is the same construct as the threshold label. The paper does include LDA topic weights and workload proxies that are not directly in the threshold rule, so the circularity is not total; however, the load-bearing narrative features are the ones used to define the outcome. The paper explicitly acknowledges in Section IV.E that the labeling heuristic 'has not yet been benchmarked against established psychometric instruments such as the Maslach Burnout Inventory,' confirming that no independent criterion validates the label. The abstract's claim of surpassing 'metadata-only baselines by ≥0.17 F1' is not backed by any displayed baseline results, and with only 43 positive providers (~4.4%), a 20% holdout contains roughly 8–9 positives, making the F1 estimate noisy. These factors converge on the same weakness: the reported 'actionable signals' are signals of the paper's own threshold rule, not of clinician burnout as a validated construct. Score 8 reflects that the central claim reduces by construction to a threshold function of its inputs, while not quite 10 because some non-threshold features (topics, workload) are included and the exact functional identity between the averaged sentiment feature and the sentence-count threshold is not fully specified.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The target variable is constructed from the same sentiment and lexicon features the model uses as predictors. The labeling thresholds, the fine-tuned BioBERT model, and the LDA settings are fitted or hand-chosen inputs. No independent psychometric benchmark is available. Workload proxies are the only features not entangled with the label, and they are not enough to ground the claim.

free parameters (4)

Burnout-label thresholds = >=12 high-confidence BioBERT sentences, >=7 stress lexicon mentions
Used to define the target variable in Section III.A; chosen without external validation or sensitivity analysis.
BioBERT sentiment fine-tuning = unspecified
The model is fine-tuned on a pseudo-labeled clinical sentiment task with no gold standard; its outputs define both features and labels.
LDA topic settings = 5 topics, 1000 Gibbs iterations, top-20 unigrams/bigrams
Model structure is chosen by coherence tuning, but no coherence metric or alternative topic counts are reported.
Positive sentiment threshold = not specified in text
Section III.A states 'sentence scores with a value of >=' but the threshold value is missing, leaving a key scoring parameter undefined.

assumptions (4)

domain assumption Discharge summaries are attributable to specific providers via Service headers and provider identifiers.
Section III.A joins notes to providers using provider IDs; if authorship attribution is wrong, all provider-level features and labels are invalid.
ad hoc to paper High-confidence negative BioBERT sentences and stress lexicon mentions are a valid proxy for burnout.
The entire ground truth is built on this assumption in Section III.A; no validation against MBI or PFI is presented.
domain assumption Narrative features reflect clinician state rather than patient acuity or documentation conventions.
The paper interprets negative sentiment and stress mentions as burnout, but they may reflect case mix, note templates, or hospital processes.
domain assumption The mapping of features to MBI dimensions is valid.
Table II aligns sentiment, pronouns, and topics to Emotional Exhaustion, Depersonalization, and Reduced Personal Accomplishment without empirical calibration.

invented entities (2)

Silver-standard burnout label (burnout proxy)
purpose: Ground truth for training and evaluating the classifier.
Defined in Section III.A as thresholds on narrative features; no independent psychometric validation is provided.
Cause-lexicon stress cues (seven burnout stressors)
purpose: Count features and a criterion for the burnout label.
A curated lexicon that is not released or validated; it is used both to define the outcome and as a predictor.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Narrative-Driven Computational Framework for Clinician Burnout Surveillance." pith.science (2026). https://pith.science/paper/ZMRBHLFX

@misc{pith2026250904497,
  author       = {Pith},
  title        = {Pith review of: A Narrative-Driven Computational Framework for Clinician Burnout Surveillance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMRBHLFX}},
  note         = {Machine review of arXiv:2509.04497}
}

read the original abstract

Clinician burnout poses a substantial threat to patient safety, particularly in high-acuity intensive care units (ICUs). Existing research predominantly relies on retrospective survey tools or broad electronic health record (EHR) metadata, often overlooking the valuable narrative information embedded in clinical notes. In this study, we analyze 10,000 ICU discharge summaries from MIMIC-IV, a publicly available database derived from the electronic health records of Beth Israel Deaconess Medical Center. The dataset encompasses diverse patient data, including vital signs, medical orders, diagnoses, procedures, treatments, and deidentified free-text clinical notes. We introduce a hybrid pipeline that combines BioBERT sentiment embeddings fine-tuned for clinical narratives, a lexical stress lexicon tailored for clinician burnout surveillance, and five-topic latent Dirichlet allocation (LDA) with workload proxies. A provider-level logistic regression classifier achieves a precision of 0.80, a recall of 0.89, and an F1 score of 0.84 on a stratified hold-out set, surpassing metadata-only baselines by greater than or equal to 0.17 F1 score. Specialty-specific analysis indicates elevated burnout risk among providers in Radiology, Psychiatry, and Neurology. Our findings demonstrate that ICU clinical narratives contain actionable signals for proactive well-being monitoring.

Figures

Figures reproduced from arXiv: 2509.04497 by the authors.

**Figure 4.** Distribution of Word Count per Provider, Showing [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 5.** Proportion of High-Severity Notes (Burnout Proxy) by [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗

Figures from the paper (1 more)

**Figure 6.** Figure 6: Five-Topic LDA Weight Profiles for the Top 10 [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

Maslach, S

C. Maslach, S. E. Jackson, and M. P. Leiter, ”Maslach Burnout Inventory Manual”, 4th ed. Menlo Park, CA: Mind Garden, 2018

work page 2018
[2]

Intensive care unit burnout: A systematic review,

M. P. Moss et al. , “Intensive care unit burnout: A systematic review,” Crit. Care Med. , vol. 44, no. 7, pp. 1414–1421, Jul. 2016

work page 2016
[3]

Burnout among health care professionals: A call to explore and address this underrecognized threat to safe, high-quality care,

L. N. Dyrbye et al., “Burnout among health care professionals: A call to explore and address this underrecognized threat to safe, high-quality care,” NAM Perspect., vol. 7, no. 7, Jul. 2017

work page 2017
[4]

Changes in burnout and satisfaction with work- life integration in physicians during the first 2 years of the COVID-19 pandemic,

T. D. Shanafelt et al., “Changes in burnout and satisfaction with work- life integration in physicians during the first 2 years of the COVID-19 pandemic,” Mayo Clin. Proc., vol. 97, no. 12, pp. 2248–2258, Dec. 2022

work page 2022
[5]

Text-mining EHR integrations for clinical decision support,

J. Adler-Milstein, T. D. Shanafelt, and A. J. Holmgren, “Text-mining EHR integrations for clinical decision support,” JAMA Netw. Open, vol. 4, no. 10, p. e2129731, Oct. 2021

work page 2021
[6]

Distributed representations of words and phrases and their compositionality,

T. Mikolov et al., “Distributed representations of words and phrases and their compositionality,” in Proc. NeurIPS, 2013, pp. 3111–3119

work page 2013
[7]

BERT: Pre-training of deep bidirectional transform- ers for language understanding,

J. Devlin et al. , “BERT: Pre-training of deep bidirectional transform- ers for language understanding,” in Proc. NAACL-HLT , 2019, pp. 4171–4186

work page 2019
[8]

MIMIC-IV , a freely accessible electronic health record dataset,

A. E. W. Johnson et al., “MIMIC-IV , a freely accessible electronic health record dataset,” Sci. Data, vol. 10, p. 219, Apr. 2023

work page 2023

Show all 19 references

[9]

Predicting physician burnout using clinical activity logs: Model performance and lessons learned,

S. S. Lou et al. , “Predicting physician burnout using clinical activity logs: Model performance and lessons learned,” J. Biomed. Inform. , vol. 127, p. 104015, Mar. 2022

2022
[10]

Predicting primary care physician burnout from electronic health record use measures,

D. Tawfik et al. , “Predicting primary care physician burnout from electronic health record use measures,” Mayo Clin. Proc. , vol. 99, no. 9, pp. 1415–1425, Sep. 2024

2024
[11]

Deep-learning model predicts physician burnout using EHR logs,

TechTarget, “Deep-learning model predicts physician burnout using EHR logs,” Aug. 24, 2022. [Online]. Available: https://www.techtarget. com

2022
[12]

Text-mining EHR integrations for clinical decision support,

J. Adler-Milstein et al. , “Text-mining EHR integrations for clinical decision support,” JAMA Netw. Open, vol. 4, no. 10, p. e2129731, Oct. 2021

2021
[13]

BurnoutEnsemble: Augmented intelli- gence to detect indications for burnout in clinical psychology,

S. Nath and M. Kurpicz-Briki, “BurnoutEnsemble: Augmented intelli- gence to detect indications for burnout in clinical psychology,” Front. Artif. Intell., vol. 4, p. 678058, Jun. 2021

2021
[14]

Design and rationale of an intelligent algorithm to detect burnout in healthcare workers in COVID era using ECG and artificial intelligence,

Anonymous, “Design and rationale of an intelligent algorithm to detect burnout in healthcare workers in COVID era using ECG and artificial intelligence,” Indian Heart J. , vol. 72, no. 6, pp. 584–588, Nov. 2020

2020
[15]

Objective measurement of physician stress in the emergency department using a wearable sensor,

E. E. Kaczor et al., “Objective measurement of physician stress in the emergency department using a wearable sensor,” in Proc. 53rd Hawaii Int. Conf. Syst. Sci. , 2020, pp. 3727–3736

2020
[16]

Using machine learning in burnout prediction: A survey,

E. E. Kaczor et al. , “Using machine learning in burnout prediction: A survey,” Pers. Ubiquitous Comput., vol. 25, no. 1, pp. 13–25, Jan. 2021

2021
[17]

AI and mental health: Predicting and preventing burnout,

A. Grig, “AI and mental health: Predicting and preventing burnout,” Medium, Jun. 6, 2024. [Online]. Available: https://medium.com

2024
[18]

AI can identify signs of staff burnout,

American Nurse Journal, “AI can identify signs of staff burnout,” Nov. 4, 2023. [Online]. Available: https://www.myamericannurse.com

2023
[19]

The role of machine learning in predicting employee burnout,

HRTech247, “The role of machine learning in predicting employee burnout,” Oct. 10, 2023. [Online]. Available: https://www.hrtech247.com

2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Maslach, S

C. Maslach, S. E. Jackson, and M. P. Leiter, ”Maslach Burnout Inventory Manual”, 4th ed. Menlo Park, CA: Mind Garden, 2018

work page 2018

[2] [2]

Intensive care unit burnout: A systematic review,

M. P. Moss et al. , “Intensive care unit burnout: A systematic review,” Crit. Care Med. , vol. 44, no. 7, pp. 1414–1421, Jul. 2016

work page 2016

[3] [3]

Burnout among health care professionals: A call to explore and address this underrecognized threat to safe, high-quality care,

L. N. Dyrbye et al., “Burnout among health care professionals: A call to explore and address this underrecognized threat to safe, high-quality care,” NAM Perspect., vol. 7, no. 7, Jul. 2017

work page 2017

[4] [4]

Changes in burnout and satisfaction with work- life integration in physicians during the first 2 years of the COVID-19 pandemic,

T. D. Shanafelt et al., “Changes in burnout and satisfaction with work- life integration in physicians during the first 2 years of the COVID-19 pandemic,” Mayo Clin. Proc., vol. 97, no. 12, pp. 2248–2258, Dec. 2022

work page 2022

[5] [5]

Text-mining EHR integrations for clinical decision support,

J. Adler-Milstein, T. D. Shanafelt, and A. J. Holmgren, “Text-mining EHR integrations for clinical decision support,” JAMA Netw. Open, vol. 4, no. 10, p. e2129731, Oct. 2021

work page 2021

[6] [6]

Distributed representations of words and phrases and their compositionality,

T. Mikolov et al., “Distributed representations of words and phrases and their compositionality,” in Proc. NeurIPS, 2013, pp. 3111–3119

work page 2013

[7] [7]

BERT: Pre-training of deep bidirectional transform- ers for language understanding,

J. Devlin et al. , “BERT: Pre-training of deep bidirectional transform- ers for language understanding,” in Proc. NAACL-HLT , 2019, pp. 4171–4186

work page 2019

[8] [8]

MIMIC-IV , a freely accessible electronic health record dataset,

A. E. W. Johnson et al., “MIMIC-IV , a freely accessible electronic health record dataset,” Sci. Data, vol. 10, p. 219, Apr. 2023

work page 2023