# Clinically-applicable prediction of hospital stay and patient similarity retrieval in paediatric cardiology using machine learning.

- **DOI:** [10.1038/s41467-026-73021-3](https://doi.org/10.1038/s41467-026-73021-3)
- **Rating:** Limited (2.8 out of 5)
- **Verdict:** Prospective NHS testing gives this paediatric cardiology pipeline practical value, but contaminated preprocessing and internally inconsistent retrieval results limit confidence.
- **Framework version:** v5.9
- **Pipeline version:** v0.4
- **Evaluation date:** 2026-10-08
- **Evaluation page:** https://www.nabu.science/paper/10.1038%2Fs41467-026-73021-3

## Key claims

1. a Random Forest model achieves the best length of stay prediction accuracy (0.88 ± 0.02), outperforming clinicians.
   - Support: Tentative
   - Type: Primary finding
   - Basis: Table 1 identifies a higher-accuracy model, while the small clinician comparison used BioClinical-BERT without a superiority test
   - Location: Abstract; Results, Table 1; Fig. 2
2. The Random Forest model shows mean precision and sensitivity of 0.77 ± 0.03 and 0.76 ± 0.04 during NHS silent deployment (1,052 admissions).
   - Support: Supported
   - Type: Secondary
   - Basis: prospective single-site monitoring supports the direction, with limited duration and a minor accuracy inconsistency
   - Location: Abstract; Results, Silent Deployment Performance; Fig. 3
3. K-means identifies three clinically distinct subgroups.
   - Support: Tentative
   - Type: Secondary
   - Basis: three clusters were recovered, but clinical separation was modest and only one LoS contrast was significant
   - Location: Abstract; Results, Patient Similarity Validation; Fig. 4; Table 4
4. Cosine similarity retrieval reveals diagnosis-driven top matches, while complications dominate broader sets.
   - Support: Tentative
   - Type: Secondary
   - Basis: Table 3 contradicts the stated retrieval pattern and contains invalid uncertainty values for bounded metrics
   - Location: Abstract; Discussion, patient-similarity paragraph; Table 3
5. In a 25-case intensive care unit pilot, clinician-rated utility improves from 4.23 ± 2.42 to 4.41 ± 2.16 (scale 0-10).
   - Support: Tentative
   - Type: Secondary
   - Basis: the uncontrolled pilot cannot establish improvement from a 0.18-point change with standard deviations above two points
   - Location: Abstract; Results, Pilot Study in Clinical Care; Table 5

## Strengths

- **Prospective monitoring in a live NHS setting** The Random Forest pipeline processed 1,052 admissions over six months without exposing predictions to clinicians. This supplies prospective performance evidence rarely available in retrospective EHR studies.
  - Dimension: contribution
  - Location: Results, Silent Deployment Performance; Fig. 3
- **Broad comparative model evaluation** Decision Tree, Random Forest, Gradient Boosting, LSTM, RoBERTa, and BioClinical-BERT variants were evaluated, with patient-level stratification and upsampling restricted to training folds.
  - Dimension: methodological rigour
  - Location: Methods, Model Comparison and Rule-based Approaches; Results, Table 1
- **Concrete clinical users and workflow** The study connects categorical LoS prediction to bed planning and tests similar-case presentation during ICU consultations, establishing a visible application pathway.
  - Dimension: application pathway
  - Location: Discussion, paragraph 2; Methods, Pilot Study in Clinical Care

## Limitations

- **Retrieval results are internally inconsistent** Table 3 contains impossible standard deviations for several bounded metrics, and its values do not support the claim that diagnosis drives top retrieval while complications dominate broader sets. This materially limits confidence in the retrieval conclusions.
  - Dimension: methodological rigour
  - Location: Results, Table 3; Discussion, patient-similarity paragraph
- **Held-out data entered model preprocessing** Training and test datasets were combined for encoding and standardisation, allowing held-out information to influence preprocessing. The prospective deployment partly tests real-world performance but does not repair the retrospective comparison design.
  - Dimension: methodological rigour
  - Location: Methods, Model Comparison and Rule-based Approaches
- **Headline comparative claims exceed evidence** Table 1 reports BioClinical-BERT accuracy of 0.89 versus 0.88 for Random Forest, while Fig. 2 compares clinicians with BioClinical-BERT on only 50 cases and reports no significance test for superiority.
  - Dimension: reporting
  - Location: Abstract; Results, Table 1; Fig. 2 caption

## Reliability checks

Caveats (4 of 4 checks).

- **Results check:** Noted
  - a Random Forest model achieves the best length of stay prediction accuracy (0.88 ± 0.02), outperforming clinicians: "The Random Forest Model performed best, demonstrating an overall accuracy of 88% across both stay categories." (Discussion, model-performance paragraph) vs "The binary classification approach yielded promising results, achieving an overall accuracy of 88% for the Random Forest Classifier and 89% for the BioClinical-BERT model." (Discussion, model-performance paragraph). Comparing the reported accuracies gives BioClinical-BERT a one-percentage-point advantage over Random Forest, contradicting the designation of Random Forest as the accuracy leader. → Not supported by the paper's own data
  - a Random Forest model achieves the best length of stay prediction accuracy (0.88 ± 0.02), outperforming clinicians: "while a Random Forest model achieves the best length of stay prediction accuracy (0.88 ± 0.02), outperforming clinicians." (Abstract) vs "Performance metrics (accuracy, precision, sensitivity, F1) of the BioClinical-BERT model, fine-tuned on the MIMIC dataset and subsequently on the GOSH dataset" (Figure 2 caption). The Abstract attributes clinician outperformance to Random Forest, whereas the caption identifying the model in the clinician comparison specifies BioClinical-BERT. → Not supported by the paper's own data
  - Cosine similarity retrieval reveals diagnosis-driven top matches, while complications dominate broader sets: "Diagnosis strongly influences early retrieval (@1), while other clinical factors,including complications, play a more prominent role at larger retrieval sets (@3, @5, @10)." (Discussion, patient-retrieval interpretation) vs "Primary diagnosis0.23 ± 0.350.25 ± 0.600.31 ± 0.310.33 ± 0.30All diagnoses0.47 ± 0.350.51 ± 0.160.53 ± 0.140.65 ± 0.60Procedures0.87 ± 0.600.85 ± 0.600.84 ± 0.580.65 ± 0.61Proportion complications0.52 ± 0.310.50 ± 0.210.48 ± 0.160.47 ± 0.13" (Table 3, diagnosis, procedure and complication rows). At @1, procedure similarity is 0.87 versus 0.23 for primary diagnosis and 0.47 for all diagnoses, while complication similarity decreases from 0.52 at @1 to 0.47 at @10 rather than becoming dominant. → Not supported by the paper's own data
  - Cosine similarity retrieval reveals diagnosis-driven top matches, while complications dominate broader sets: "Primary diagnosis0.23 ± 0.350.25 ± 0.600.31 ± 0.310.33 ± 0.30All diagnoses0.47 ± 0.350.51 ± 0.160.53 ± 0.140.65 ± 0.60Procedures0.87 ± 0.600.85 ± 0.600.84 ± 0.580.65 ± 0.61" (Table 3). For observations bounded by 0 and 1, sample SD is at most sqrt(n·mean·(1−mean)/(n−1)); for n≥3 this bounds the flagged primary-diagnosis, all-diagnoses and first three procedure SDs by 0.530, 0.584, 0.412, 0.437 and 0.449 respectively, and n=2 permits still smaller maxima, excluding the reported 0.60, 0.60, 0.60, 0.60 and 0.58. → Effect on result unclear
  - a Random Forest model achieves the best length of stay prediction accuracy (0.88 ± 0.02), outperforming clinicians: "BioClinical BERT (FT-MIMIC-GOSH)0.77 ± 0.140.94 ± 0.020.82 ± 0.050.79 ± 0.090.94 ± 0.020.77 ± 0.140.92 ± 0.930.93 ± 0.020.89 ± 0.04" (Table 1, BioClinical BERT (FT-MIMIC-GOSH), long-stay precision). A sample precision mean of 0.92 on a bounded 0–1 measure permits SD no greater than sqrt(2×0.92×0.08)=0.384 for any sample size of at least two, so the printed SD of 0.93 is impossible. → Effect on result unclear
  - The Random Forest model shows mean precision and sensitivity of 0.77 ± 0.03 and 0.76 ± 0.04 during NHS silent deployment (1,052 admissions): "During the silent mode deployment, the model achieved a mean accuracy of 0.776 (±0.021)." (Results, Silent deployment performance) vs "It demonstrated consistent performance, maintaining a high overall accuracy of 0.766 ± 0.021." (Discussion, silent deployment). The two reported mean accuracies for the same deployment differ by 0.776−0.766=0.010, and the supplied text does not identify which is correct. → Effect on result unclear
  - Minor reporting notes (1):
    - Descriptive: "Gradient boosting0.76 ± 0.080.90 ± 0.020.76 ± 0.050.76 ± 0.060.90 ± 0.020.90 ± 0.020.90 ± 0.030.90 ± 0.030.86 ± 0.04" (Table 1, Gradient boosting row) vs "LoSShort stayLong staySensitivitySpecificityPrecisionF1 scoreSensitivitySpecificityPrecisionF1 scoreAccuracy" (Table 1, column headings). In a binary classification evaluated on the same predictions, long-stay specificity must equal short-stay sensitivity, but the Gradient Boosting row reports 0.90 ± 0.02 and 0.76 ± 0.08 respectively. → Changes the size materially
- **Reporting and transparency:** Noted. Post-hoc testing, pilot feedback denominators and the pilot rating scale need clarification; the deployment end date and procedure-coding label contain correctable reporting errors.
  - Ethics approval: declared
- **Reference integrity:** Clean. 61 of 61 checkable references verified
  - 77 references in manuscript
  - 16 have no canonical index record — counted, but not index-checkable
  - 2 references confirmed by manual review
  - References not verified automatically are resolved manually within 24 hours.

## Post-publication record

- **Status:** Clean
- **Summary:** No retraction or expression of concern found in Crossref (includes Retraction Watch).
- **Last checked:** 2026-10-17
