An electronic frailty index built by deep learning on clinical notes: across 193,629 Finns, a sevenfold death risk in the frailest — but a threshold calibrated on the mortality it predicts

A team led from Tampere University builds a next-generation electronic frailty index: on top of coded diagnoses and lab tests, a deep-learning model extracts ten functional deficits (mobility, vision, hearing, isolation, need for help) from free clinical notes. Across 193,629 Finns aged 35 to 103, the death risk of the frailest is multiplied by 7.3 and severe-infection risk by 9.2. But the thresholds that define frailty are calibrated on the mortality the index then predicts, the only comparators are two code-based scores used outside their setting, and nothing is validated beyond a single Finnish county.

The context

Frailty is a state of increased vulnerability linked to declining physiological reserves and a reduced capacity to return to equilibrium after a stress. It strongly predicts mortality, hospitalisation and admission to long-term care, yet it is rarely measured in routine practice. The idea of a frailty index (FI) is to count the "health deficits" accumulated across the body's systems: the more you accumulate, the frailer you are. Ported into the electronic health record, this becomes an electronic frailty index (eFI).

The problem the paper tackles: existing eFIs are mostly calibrated on older people, rely on structured data (diagnostic codes) and a single care setting. The UK eFI, the most widely deployed, rests on Read Codes, a coding system specific to the NHS and hard to transfer elsewhere, where diagnoses are usually coded with ICD (International Classification of Diseases). Above all, structured data capture poorly what makes frailty in daily life — mobility, incontinence, isolation, vision, hearing — because that information lives in the free text of clinical notes, not in a list of codes.

The method

The team, based at Tampere University (Finland) with co-authors in Italy, Sweden, Australia and Hong Kong, builds its eFI from the entire records of a Finnish health authority — the Central Finland county, about 270,000 inhabitants — over 2010–2023. The analytic population is 193,629 people born between 1920 and 1975, aged 35 to 103 (mean age 62.0 years, standard deviation 13.6; 51.6% women).

The index aggregates 53 equally weighted items, each coded present or absent: 34 ICD-10 diagnoses, 9 laboratory results, and 10 deficits extracted from free text by a deep-learning NLP (natural language processing) model — falls, incontinence, loneliness, mobility limitations, hearing impairment, visual impairment, age-related neurocognitive problems, and need for help with bathing, dressing and eating. Named-entity recognition (NER, the automatic spotting of these notions in text) reaches F1 scores (a mean of precision and recall, from 0 to 1) of 0.74 to 0.92 depending on the item. The eFI is rescaled from 0 to 100 (median 2.3). Data-driven cut points (8, 15, 21) carve out four categories: non-frail, mild, moderate, severe.

The predicted outcomes are all-cause mortality (Cox models, which estimate a relative risk over time), severe infections requiring hospitalisation and fractures (at 2 years), and healthcare use. The comparators are the Hospital Frailty Risk Score (HFRS) and the Charlson Comorbidity Index (CCI), both code-based. Validation rests on an internal 70/30 (training/test) split of the same dataset.

The results

In this population, 77.0% are classed non-frail, 14.4% mild, 5.9% moderate and 2.7% severe (5,260 people). The associations are strong and graded. For mortality, the relative risk (hazard ratio, HR) of the severely frail versus non-frail reaches 7.31 (95% confidence interval: 6.83–7.83); it is 2.14 for mild and 3.97 for moderate. For severe infections, the HR of the severe group climbs to 9.22 (8.52–9.98) — well beyond the HR of 3.37 reported in the frailest category of a UK Biobank study cited for comparison. For fractures, HR 2.75 (2.52–3.01), with a C-index (the probability of correctly ranking two patients by risk) of 0.62, comparable to FRAX (~0.61), the reference tool for fracture risk. Continuously, each eFI point raises the death risk by 9% (HR 1.09).

A point the authors emphasise: frailty is already present in midlife — severe frailty in 2.0% of 35–49-year-olds and 11.6% of 50–64-year-olds, versus 35.1% of 65–79-year-olds and 51.3% of those 80 and over. This is the central argument of the title, "the full frailty spectrum." The authors state that the eFI brings "the greatest improvement in discrimination" across all endpoints, ahead of HFRS and CCI. One result jars, however: for the probability of using care, the odds ratio of the severe group (3.15) is lower than that of the moderate group (3.56) — a non-monotonicity the paper does not discuss.

What is good

Reading what codes discard. The real contribution is extracting from free text ten functional and sensory deficits — mobility, vision, hearing, isolation, need for help with daily activities — that eFIs built on structured codes almost always miss. The NLP holds up (F1 of 0.74 to 0.92), and these items are precisely the ones that give the notion of frailty its clinical meaning.

An uncommon scale and age window. 193,629 people followed from 2010 to 2023, from age 35, in a public system with universal coverage that limits insurance-related selection bias. Extending the measurement of frailty into midlife, and showing it is already present there and associated with outcomes, is a real contribution: most eFIs stop at those over 65.

Honest clinical anchoring. The authors tie some figures to verifiable benchmarks — the fracture C-index of 0.62 is explicitly set against FRAX (~0.61) — and show a coherent dose-response gradation across categories. The index construction follows established methodological criteria for frailty indices, making it comparable to the literature.

What is less good

A threshold calibrated on what it predicts. This is the central limitation, and it is data leakage through circularity. The category boundaries (8, 15, 21) are set "from the observed mortality risk" in the dataset, then the same data are used to measure the association between eFI and mortality. The cut-off rule is calibrated on the very target it then aims to predict, which mechanically inflates already spectacular HRs (7.31 for mortality, 9.22 for infections). A mere 70/30 split does not cancel this bias if the thresholds are derived on the whole sample.

A disadvantaged comparator, and no eFI-versus-eFI comparison. This is a biased comparator. The whole paper is built against the limits of the UK eFI — yet it never tests it. The only comparators are the HFRS, a score designed for hospitalised older patients and applied here to an entire 35-to-103 population, and Charlson, a plain comorbidity score. "Beating" two code-based scores used outside their original setting does not show you beat an established eFI.

The numbers that matter are missing, and validation is single-site. This is both a misleading metric and a population bias. The body of the paper gives no AUROC, no calibration curve, no event count: the "greatest improvement in discrimination" remains a qualitative claim, the figures deferred to appendices. Validation is internal (70/30 split), retrospective, on a single Finnish county — a homogeneous, almost exclusively white population, a Nordic care system — with no external cohort. Transportability to other countries and coding systems is asserted but never demonstrated. Finally, the NLP model is not named, no code is shared, the data are under restricted access, the licence is non-commercial and no-derivatives (CC BY-NC-ND), and this is a non-peer-reviewed preprint.

What it changes

For the research community, the useful signal is not "an HR of 9," it is "free text adds functional deficits that codes do not see." That is reusable — provided the model is named, the code published, and above all the tool validated externally. The methodological lesson is clear: when risk categories are cut on the very outcome you will then predict, the reported hazard ratios overestimate real-world value; the only credible judge would be an external cohort, with AUROC and calibration shown in plain sight.

For clinicians, nothing is deployable today: a single-centre research tool, not prospectively validated, unapproved, and designed for a Finnish coding system. The idea of a frailty dashboard that reads the notes so as to lose nothing of the patient's daily life is credible in the medium term, in support of the geriatrician, not in their place. For patients and the general public, the nuance matters: a "sevenfold death risk" describes a population carved out from mortality itself, not a prediction validated elsewhere. Which takes nothing away from the value of spotting earlier, from one's fifties, a frailty that remains widely under-diagnosed.

To go further

The preprint is available on medRxiv (10.64898/2026.06.23.26356323). On extracting information from clinical notes with language models, see our decryptage of Genosolver, which reads reports to diagnose rare diseases; on the external validation of a risk model and its performance drop away from its home centre, that of a 10-year stroke risk model; and on the gap between headline performance and real-world validation, that of a real-world validation of an adherence model.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.