Suboptimal Echocardiographic Image Quality Predicts in‐Hospital Mortality Across the Adult Age Spectrum, Independent of Admission Type: A Large‐Scale NLP Study

ABSTRACT

Purpose

Suboptimal echocardiographic image quality is common and usually treated as a reason for study exclusion, yet discarded studies come disproportionately from older patients. We examined whether it independently predicts in‐hospital mortality across the adult age spectrum.

Methods

We analyzed 32,813 reports from 19,731 patients (ECHO‐NOTE2NUM/MIMIC‐III, 2001–2012). A validated natural language processing (NLP) classifier (κ = 0.996) identified suboptimal studies. Multivariable logistic regression, GEE, and IPTW assessed mortality across five age groups (18–49, 50–64, 65–74, 75–84, ≥85 years).

Results

Overall, 9,150 studies (27.9%) were suboptimal, with similar rates across age groups but differing causes (tachycardia and body habitus in younger versus poor acoustic windows in older patients). Suboptimal imaging independently predicted mortality (IPTW OR 1.47, 95% CI 1.41–1.54). Age modified this association (interaction p = 0.04): the effect was consistently elevated across 18–74 and attenuated to 1.20 (1.06–1.35) in patients aged ≥85. Comorbidity adjustment only partly attenuated the age gradient (Charlson‐adjusted interaction p = 0.12), while documented goals‐of‐care limitations (Do Not Resuscitate / Do Not Intubate; DNR/DNI) rose from 7.9% (18–49) to 32.0% (≥85), peaking in the same stratum in which the mortality association disappeared.

Conclusion

Suboptimal image quality independently predicts in‐hospital mortality across a broad 18–74 vulnerability window, with attenuation in patients aged ≥85 that is only partly explained by comorbidity and parallels a steep rise in goals‐of‐care limitations. Image quality is best read as an age‐dependent prognostic signal whose actionability declines as care shifts toward comfort in the oldest old, supporting active escalation in adults aged 18–74 and goals‐concordant interpretation thereafter.

In 32,813 echocardiograms, suboptimal image quality independently predicted in‐hospital mortality (OR 1.47), with the association strong through age 74 and attenuating beyond 85 as goals of care shifted. Rather than a reason to discard a study, image quality is best read as an age‐dependent prognostic signal.

1. Introduction

As populations age, a growing share of diagnostic testing is performed in older adults whose anatomy, comorbidity burden, and goals of care complicate both the acquisition and the interpretation of results. Echocardiography is the most widely used cardiac imaging modality in this population, valued for its portability, absence of ionizing radiation, and applicability from the outpatient clinic to the intensive care unit (ICU) [1, 2]. Yet adequate image quality cannot be achieved in every patient even under guideline‐concordant practice [1, 3], and the patients in whom imaging is hardest to obtain are disproportionately older and sicker. How the diagnostic limitations of imaging translate into outcomes across the adult age spectrum, and whether their prognostic meaning shifts with advancing age, is a question of direct relevance to geriatric medicine that has not been quantitatively addressed.

Yet its clinical consequences remain poorly characterized. Reported rates range from 10%–30% in general populations to nearly 50% in intensive care, where mechanical ventilation, limited repositioning, body habitus, and lung pathology degrade ultrasound penetration [4, 5]. What follows a suboptimal study—whether a missed or delayed diagnosis ensues, and whether outcomes differ—has attracted surprisingly little attention.

The relationship between patient age and the consequences of suboptimal imaging has received even less attention, though the rationale is straightforward. Older patients carry more cardiac pathology to each encounter (valvular degeneration, diastolic dysfunction, left ventricular hypertrophy), and therefore more that may go undetected when image quality is inadequate [6, 7], while age‐related thoracic changes (kyphosis, increased anteroposterior diameter, emphysematous lung fields) independently compromise acoustic windows [8]. Decision‐making in older adults is further complicated by multimorbidity, frailty, and competing priorities, such that diagnostic uncertainty carries greater consequences than in younger patients; from a geriatric perspective, a test's value depends not only on technical adequacy but on how its results align with functional status, life expectancy, and goals of care [8, 9].

Whether these converging factors produce measurably different reporting behavior, and whether a suboptimal study carries different prognostic significance with age, remains unanswered. If the mortality impact concentrates in a specific age window, quality benchmarks, contrast protocols, and referral thresholds could be calibrated accordingly.

Echocardiography reports contain a dimension of diagnostic information seldom formally analyzed: the hedging language used when image quality or clinical complexity prevents definitive interpretation. Phrases such as “cannot be excluded” and “unable to assess” carry direct clinical weight, prompting additional testing, delaying treatment, or leaving the referring team without a definitive answer [10]. Natural language processing (NLP) enables extraction of this language across large corpora [11, 12]. The EchoGraph model, a BERT‐based system trained on echocardiography reports, classifies each observation as definitely present, definitely absent, or uncertain, enabling structured characterization beyond keyword matching [13, 14].

No prior work has examined image quality as a prognostic variable at this scale or across the full adult age spectrum. We analyzed 32,813 echocardiography reports from 19,731 patients in a large publicly available critical care database, with validated ICU severity scores (SAPS‐II and OASIS) linked to 99.6% of the cohort, pursuing four aims: to map age‐specific prevalence and causes of suboptimal imaging; to quantify how diagnostic uncertainty language varies with age; to test whether suboptimal imaging is independently associated with in‐hospital mortality; and to determine whether this association differs across the age spectrum.

2. Materials and Methods

2.1. Data Source and Study Population

We conducted a retrospective cohort analysis of the EchoGraph‐annotated ECHO‐NOTE2NUM dataset [13, 14, 15] (45,794 reports), derived from MIMIC‐III v1.4 (Beth Israel Deaconess Medical Center, Boston, 2001–2012) [16, 17]. Access was obtained via standard PhysioNet credentialing.

Reports were linked to the MIMIC‐III PATIENTS table (sex, date of birth, vital status) and ADMISSIONS table (admission date, admission type, insurance, ethnicity, in‐hospital mortality) via subject and admission identifiers. Age at echocardiography was calculated from date of birth and admission date; per MIMIC‐III protocols, patients older than 89 years were assigned an age of 90 years, which was handled explicitly in analyses.

We retained reports from adult patients (≥18 years) with complete demographic linkage to both tables. A total of 12,981 reports without admission linkage (predominantly outpatient echocardiograms) were excluded, leaving 32,813 selected reports from 19,731 patients across 22,913 admissions (Figure S1).

Patients were stratified into five clinically meaningful age groups: 18–49 years (young adults), 50–64 years (middle‐aged adults), 65–74 years (young‐old), 75–84 years (old‐old), and ≥85 years (oldest‐old). These strata follow established geriatric classifications that distinguish the young‐old, old‐old, and oldest‐old, reflecting clinically meaningful differences in functional status, comorbidity burden, and goals of care across the older adult population, and aligning with frameworks recognizing distinct age‐related phenotypes in cardiovascular care [8, 9].

2.2. NLP Classification of Suboptimal Image Quality

A keyword‐based NLP classifier identified suboptimal studies from free‐text interpretations, flagging reports containing any of eight ASE‐standard phrases [1, 18] (“suboptimal,” “poor acoustic windows,” “poor echo windows,” “limited study,” “technically difficult,” “difficult to image,” “body habitus,” “inadequate windows”). Validation against 500 randomly selected reports used an expanded linguistic pattern set; discordant cases were adjudicated by two investigators, with disagreements resolved by consensus. Performance was near‐perfect (Table S10): sensitivity 100% (219/219), specificity 99.6% (280/281), positive predictive value 99.5% (219/220), negative predictive value 100% (280/280), overall accuracy 99.8%, and Cohen's κ = 0.996.

2.3. Diagnostic Uncertainty Language Extraction

We searched reports for eight established phrases signaling diagnostic uncertainty (including “cannot be excluded,” “cannot assess,” “unable to assess,” “not well visualized,” and “indeterminate”), each recorded as binary per report. Three low‐prevalence terms with substantial semantic overlap were excluded from the composite “any uncertainty” variable, defined as presence of at least one of five clinically distinct phrases. Prevalence of each was compared across age groups and between suboptimal and optimal studies.

2.4. EchoGraph Structured Entity Analysis

Reports were annotated using the EchoGraph model [13], a BERT‐based information extraction system (entity F1 0.85 internal, 0.80 external) that classifies elements as Observation‐Definitely Present (OBS‐DP), Observation‐Definitely Absent (OBS‐DA), Anatomy, or Measurements. For each report, entity counts were tabulated and a positivity ratio calculated as OBS‐DP / (OBS‐DP + OBS‐DA), excluding zero‐observation reports, capturing whether older patients’ reports are weighted toward confirmed pathology over excluded findings.

2.5. Statistical Analysis

This study follows STROBE guidelines [19]; a completed STROBE checklist is provided as Supplementary File S1. Missing data were negligible (<1%) and handled by complete‐case analysis. The Charlson comorbidity index (derived from ICD‐9 codes) and code‐status‐based goals‐of‐care variables were available for all admissions; SAPS‐II and OASIS were available for 99.6% of reports, with the remaining 0.4% corresponding to admissions without a linked ICU stay, and missingness was non‐differential across image‐quality groups. Categorical variables were compared by chi‐square; continuous variables (mean ± SD) by Kruskal‐Wallis given non‐normal entity count distributions.

Multivariable logistic regression assessed two primary outcomes: (1) “cannot be excluded” uncertainty language and (2) in‐hospital mortality. For mortality, three sequential models were built: Model A (suboptimal status only), Model B (added age per 10‐year increment, sex, emergency admission), and Model C (added NLP‐derived OBS‐DP, OBS‐DA, measurement count, and “cannot be excluded” status). Discrimination was assessed by C‐statistic and fit by AIC.

The age‐by‐suboptimal interaction on mortality was tested with an interaction term in the logistic regression model, and its significance assessed by likelihood ratio test comparing nested models. Stratum‐specific odds ratios were then calculated within each of the five age groups to characterize the gradient.

Confounding by indication was addressed using inverse probability of treatment weighting (IPTW) [20] with propensity scores estimated from age, sex, and admission type; clinical confounders not directly captured were addressed in a subsequent severity‐adjusted sensitivity analysis. Weights were trimmed at the first and 99th percentiles. The primary analytic unit was the echocardiography report (n = 32,813 from 19,731 unique patients across 22,913 admissions); because 35.8% of patients contributed more than one report (median 1, range 1–34), within‐patient clustering (mean 1.66 reports per patient) was handled using generalized estimating equations (GEE) with an exchangeable correlation structure as a co‐primary analysis, and a pre‐specified first‐echocardiogram‐per‐patient analysis (n = 19,731) was provided as a sensitivity check [21]. Covariate balance was quantified with standardized mean differences (Table S2, Figure S2), sensitivity to unmeasured confounding with E‐values [22] (Table S3), and calibration with the Hosmer‐Lemeshow test (Figure S3); formal interaction terms tested age‐by‐suboptimal and sex‐by‐suboptimal effects on mortality.

To support the geriatric interpretation, we linked the cohort to two additional domains. Comorbidity burden was quantified using the Charlson index from ICD‐9 codes (Quan adaptation [23]) with the age component removed, so comorbidity and chronological age could be modeled independently. Goals‐of‐care intensity was captured from documented code‐status orders, flagging each admission for any limitation (do‐not‐resuscitate/do‐not‐intubate) and, separately, comfort‐measures‐only status. We then (i) repeated the age‐stratified mortality models with comorbidity adjustment to test whether the interaction was independent of comorbidity burden, and (ii) examined code‐status limitations across age strata to test whether attenuation of the mortality association in the oldest patients coincided with a shift toward goals‐of‐care‐directed management.

Pre‐specified sensitivity analyses included restriction to the first echocardiogram per patient, exclusion of age‐capped patients, temporal cohort analysis, and gradient boosting machine models (Table S6). To address residual confounding by illness severity, a post‐hoc analysis used two validated ICU scores from the MIMIC‐III concepts repository [16], SAPS‐II [24] and OASIS [25] (worst score per admission), fitted overall and within each age stratum. All analyses used Python 3.12. Two‐sided p < 0.05 was considered significant; given the large sample, we emphasize effect sizes over p‐values.

3. Results

3.1. Study Population Characteristics

The analytic cohort included 32,813 echocardiography reports from 19,731 unique patients (Table 1). Age distribution reflected the critical care population captured by MIMIC‐III, ranging from 15.4% in 18–49‐year‐olds to 9.9% in patients aged ≥85. The proportion of female patients was lowest at 50–64 and rose thereafter (40.6% in the youngest to 57.3% in the oldest group; p < 0.001), consistent with the sex differential in longevity. Emergency admissions comprised the majority across all groups (highest in the youngest and oldest strata). In‐hospital mortality increased progressively from 11.5% in 18–49‐year‐olds to 20.1% in patients aged ≥85.

TABLE 1.

Baseline characteristics (N = 32,813 reports, 19,731 patients). Insurance and race/ethnicity categories shown are the most frequent and are not exhaustive; remaining patients fall into other or unspecified categories. Percentages are column‐wise within each age group.

Characteristic	18–49 (n = 5,054)	50–64 (n = 9,451)	65–74 (n = 7,577)	75–84 (n = 7,475)	≥85 (n = 3,256)	p
Reports, N (%)	5,054 (15.4)	9,451 (28.8)	7,577 (23.1)	7,475 (22.8)	3,256 (9.9)	—
Patients, N (%)	2,830 (14.3)	5,526 (28.0)	4,604 (23.3)	4,814 (24.4)	2,303 (11.7)	—
Age (mean ± SD)	39.3±8.4	57.7±4.3	69.5±2.8	79.3±2.8	88.0±2.0	<0.001
Female, N (%)	2,054 (40.6)	3,408 (36.1)	3,070 (40.5)	3,472 (46.4)	1,867 (57.3)	<0.001
Race/ethnicity
White	3,123 (61.8)	6,746 (71.4)	5,526 (72.9)	5,641 (75.5)	2,654 (81.5)	<0.001
Black	645 (12.8)	806 (8.5)	554 (7.3)	360 (4.8)	134 (4.1)	<0.001
Other/unknown	1,286 (25.4)	1,899 (20.1)	1,497 (19.8)	1,474 (19.7)	468 (14.4)	<0.001
Insurance
Medicare	904 (17.9)	2,006 (21.2)	6,239 (82.3)	7,018 (93.9)	3,140 (96.4)	<0.001
Private	2,490 (49.3)	4,428 (46.9)	884 (11.7)	250 (3.3)	47 (1.4)	<0.001
Admission type
Emergency	4,406 (87.2)	7,735 (81.8)	6,134 (81.0)	6,172 (82.6)	2,901 (89.1)	<0.001
Elective	490 (9.7)	1,410 (14.9)	1,228 (16.2)	1,074 (14.4)	301 (9.2)	<0.001
Mortality, N (%)	579 (11.5)	996 (10.5)	963 (12.7)	1,209 (16.2)	656 (20.1)	<0.001

Age capped at 90 per MIMIC‐III. Patient counts may sum >19,731 (multiple admissions across age groups). Mortality reflects report‐level in‐hospital death and is not additive across strata.

Not mutually exclusive. % among suboptimal.

C‐stat: A = 0.546, B = 0.611, C = 0.614.

Among 32,813 reports, 27,348 (83.3%) were performed during emergency admissions, 962 (2.9%) during urgent admissions, and 4,503 (13.7%) during elective admissions, with no missing admission type. The proportion of emergency admissions remained stable across age groups (range 81.0%–89.1%), with the oldest patients (≥85 years) most likely to be admitted emergently (89.1%). In‐hospital mortality was 14.7% in emergency admissions, 12.1% in urgent admissions, and 5.4% in elective admissions.

3.2. Prevalence and Age‐Specific Causes of Suboptimal Imaging

Overall, 9,150 echocardiograms (27.9%) met criteria for suboptimal image quality (Table S11, Figure 1A; age‐specific causes shown in Figure 1B). The suboptimal rate was lowest in the youngest group (24.2% in 18–49 years) and relatively uniform across older groups (28.4–28.7%; overall p < 0.001). Although cardiovascular comorbidities are more prevalent in older patients, suboptimal rates did not increase monotonically with age; the pattern reflected a shift in dominant causes rather than a rising overall frequency.

FIGURE 1.

Age‐stratified analysis of echocardiographic image quality and clinical outcomes (N = 32,813). (A) Suboptimal imaging rate by age group. (B) Stacked proportions of causes underlying suboptimal imaging. (C) Diagnostic uncertainty language by age group. (D) IPTW‐adjusted odds ratios for mortality by age group (p‐interaction = 0.04). (E) Positivity ratio by age group. (F) In‐hospital mortality stratified by image quality and age group.

Behind these comparable overall figures lay markedly different causes. Tachycardia was nearly three times more frequent in the youngest vs. oldest group (14.8% vs. 5.2%; p < 0.001), body habitus 2.7 times more common in younger patients (16.3% vs. 6.1%; p < 0.001) [26], and mechanical ventilation also more frequent in the young (16.1% vs. 10.4%; p < 0.001). By contrast, poor acoustic windows, the most common cause overall, predominated across all age groups (49%–58%), peaking at 58.4% in the 65–74 group. Patient positioning difficulty did not differ significantly across age (11.5%–14.2%; p = 0.126).

The similar overall rates mask different underlying mechanisms: in younger patients the limiting factors are acute and physiological (tachycardia, ventilator settings, body habitus), whereas in older patients they are predominantly structural and anatomical.

3.3. Age‐Related Patterns in Diagnostic Uncertainty Language

The phrase “cannot be excluded” represents the most clinically consequential form of diagnostic uncertainty, as it implies potential pathology that the interpreting physician is unable to confirm or refute. Its prevalence rose steadily and monotonically across age groups (3.5% at 18–49 years, 5.0% at 50–64, 6.2% at 65–74, 6.2% at 75–84, and 6.7% at ≥85 years; p < 0.001; Table S7, Figure 1C), a near‐doubling from the youngest to the oldest group.

In multivariable logistic regression, age remained independently associated with “cannot be excluded” language after adjustment for sex, suboptimal status, and admission type (odds ratio [OR] 1.13 per 10‐year increment, 95% CI 1.10–1.17; p < 0.001). Suboptimal image quality was the strongest predictor (OR 4.42, 95% CI 4.01–4.88; p < 0.001), followed by female sex (OR 1.35, 95% CI 1.23–1.49; p < 0.001) and emergency admission (OR 1.53, 95% CI 1.31–1.79; p < 0.001). The age‐by‐suboptimal interaction for uncertainty language was not statistically significant (p = 0.32), indicating that suboptimal imaging amplifies diagnostic uncertainty to a similar degree regardless of patient age.

The term “indeterminate,” by contrast, declined from 9.6% to 5.5% (p < 0.001). The two phrases occupy different niches: “indeterminate” reflects a Doppler‐dependent measurement that cannot be obtained, whereas “cannot be excluded” signals uncertainty about a structural finding requiring adequate 2D visualization. Their divergent age trajectories reflect which challenge dominates as patients age.

When stratified by both age and image quality (Table S8), “cannot be excluded” language was consistently 3‐ to 4‐fold more prevalent in suboptimal than optimal studies across every age group. Even in technically adequate studies the rate doubled across the age spectrum (2.0% at 18–49 to 4.1% at ≥85), indicating that age contributes its own layer of diagnostic complexity, independent of image quality.

3.4. EchoGraph Entity Analysis: An Age‐Related Shift in Reporting Patterns

EchoGraph entity analysis revealed a progressive shift in reporting patterns with age (Table S9, Figure S4). Definitely absent observations decreased monotonically (2.1 to 1.4 per report; p < 0.001), while definitely present observations showed a U‐shaped pattern. The positivity ratio increased linearly from 0.820 to 0.870 across age groups (p < 0.001; Figure 1E): in the youngest patients roughly one‐fifth of observations excluded pathology, versus one‐seventh in the oldest. Measurement count decreased from 4.7 to a nadir of 3.7 at 65–74 before partially recovering, probably reflecting differences in study thoroughness.

3.5. Suboptimal Imaging as an Independent Predictor of in‐Hospital Mortality

Suboptimal imaging was independently associated with in‐hospital mortality, with the strength of this association varying by patient age (Tables 2 and 3; Figure 1F).

TABLE 2.

Age‐stratified suboptimal‐mortality association.

Age	N	Unadj OR	95% CI	IPTW OR	95% CI	p
18–49	5,054	1.615	1.339–1.947	1.599	1.418–1.803	<0.001
50–64	9,451	1.645	1.436–1.886	1.608	1.469–1.760	<0.001
65–74	7,577	1.707	1.483–1.964	1.637	1.490–1.797	<0.001
75–84	7,475	1.430	1.254–1.630	1.385	1.271–1.509	<0.001
≥85	3,256	1.233	1.024–1.484	1.199	1.062–1.353	0.003

Interaction: OR 0.953/10 yr (0.910–0.998; p = 0.04).

Post‐hoc severity‐adjusted odds ratios for the suboptimal–mortality association in 32,682 echocardiograms (99.6% of evaluable studies). The age‐stratified pattern is preserved after severity adjustment, with effects remaining elevated across all strata below age 85 and significant attenuation in patients aged ≥85. The severity subset comprised 99.6% of the evaluable cohort.

TABLE 3.

Comorbidity‐adjusted mortality association and goals‐of‐care orders by age group.

Age group (years)	Reports (n)	Charlson‐adjusted OR (95% CI) a	Any code‐status limitation (%) b	Comfort‐measures‐only (%)
18–49	5,054	1.68 (1.39–2.04)	7.9	2.7
50–64	9,451	1.55 (1.35–1.78)	8.2	2.9
65–74	7,577	1.65 (1.43–1.90)	11.2	3.3
75–84	7,475	1.42 (1.24–1.61)	18.5	5.3
≥85	3,256	1.24 (1.03–1.49)	32.0	6.0

Odds ratios are for the association between suboptimal image quality and in‐hospital mortality within each age stratum, derived from logistic regression adjusted for the Charlson comorbidity index with its age component removed (n = 32,813). These differ modestly from the IPTW‐weighted estimates in Table 2 because of the different adjustment method (comorbidity regression versus propensity weighting on administrative covariates); the two approaches yield a similar age pattern (same rank order of strata, though the interaction reached significance only in the unadjusted/IPTW analysis).

^a The age‐by‐suboptimal interaction was attenuated after comorbidity adjustment (likelihood‐ratio p = 0.12); the overall comorbidity‐adjusted OR was 1.50 (vs. 1.53 unadjusted), with a Charlson OR of 1.20 per point.

^b Any code‐status limitation denotes a documented do‐not‐resuscitate or do‐not‐intubate order during the admission. Code‐status limitations rose roughly fourfold across the age spectrum, peaking in the same stratum (≥85) in which the mortality association lost significance, supporting a goals‐of‐care interpretation of the attenuation.

In the unadjusted analysis (Model A), suboptimal imaging was associated with a 53% increase in the odds of in‐hospital mortality (OR 1.53, 95% CI 1.43–1.64; p < 0.001). After adjustment for age, sex, and admission type (Model B; Table S12), the association remained highly significant (OR 1.47, 95% CI 1.38–1.58; p < 0.001; C‐statistic 0.611). IPTW analysis yielded a consistent estimate (OR 1.47, 95% CI 1.41–1.54; p < 0.001) with excellent covariate balance (all SMD <0.01; Table S2). GEE analysis accounting for within‐patient clustering (mean 1.66 reports per patient; exchangeable correlation 0.72) yielded an attenuated but still significant estimate (OR 1.23, 95% CI 1.16–1.31; p < 0.001; Table S1), indicating that the association is not an artifact of non‐independence. E‐value analysis showed that an unmeasured confounder would need a risk ratio of at least 2.02 with both exposure and outcome to account for even the lower confidence bound of the IPTW estimate (Table S3).

Adding NLP‐derived entity features (Model C) improved discrimination modestly but significantly (C‐statistic 0.614 vs. 0.611). Each additional confirmed pathological finding carried slightly higher mortality odds (OR 1.021 per entity; p < 0.001), consistent with sicker patients accumulating more documented disease, while measurement count was inversely associated with mortality (OR 0.960 per entity; p < 0.001), likely reflecting study thoroughness.

3.6. Age‐Dependent Association between Suboptimal Imaging and Mortality

The age‐by‐suboptimal interaction on mortality was statistically significant (likelihood ratio test p = 0.04; Table 2; Figures 1D and 2), with a gradient that tracked the intensity of active therapeutic decision‐making across age groups.

FIGURE 2.

Forest plot of IPTW‐adjusted odds ratios for in‐hospital mortality (orange) and “cannot be excluded” uncertainty language (blue) by age group. Diamond indicates overall IPTW estimate (OR 1.47; 95% CI 1.41–1.54). Figure S5. Comprehensive sensitivity analysis forest plot. Odds ratios (95% CI) for the suboptimal‐mortality association across all pre‐specified and post‐hoc sensitivity analyses. Blue squares: primary and pre‐specified sensitivity analyses. Gold squares: post‐hoc severity‐adjusted analyses (SAPS‐II and OASIS, n = 32,682). The red dashed line at OR = 1 indicates no association. The age‐stratified pattern, including the broad 18–74 vulnerability window, is preserved across all severity‐adjusted analyses.

In the youngest patients (18–49 years), suboptimal imaging was associated with significantly increased mortality (unadjusted OR 1.62, 95% CI 1.34–1.95; IPTW OR 1.60, 95% CI 1.42–1.80). This association strengthened slightly in the 50–64 age group (OR 1.65; IPTW OR 1.61) and remained elevated in the 65–74 “young‐old” group (OR 1.71, 95% CI 1.48–1.96; IPTW OR 1.64, 95% CI 1.49–1.80), with confidence intervals overlapping across the 18–74 range. From age 75 onward, the association progressively attenuated: OR 1.43 (95% CI 1.25–1.63; IPTW OR 1.39) in the 75–84 group, and OR 1.23 (95% CI 1.02–1.48; IPTW OR 1.20, 95% CI 1.06–1.35) in those aged ≥85. In the interaction model, each 10‐year age increment reduced the suboptimal‐mortality OR by a factor of 0.953 (95% CI 0.910–0.998).

The association held across all pre‐specified sensitivity analyses (Table S6; Figure S5): first echocardiogram per patient (OR 1.56, 1.43–1.70), exclusion of age‐capped patients (OR 1.50, 1.40–1.60), and temporal halves (period interaction p = 0.47; Table S4). Restricting the analysis to emergency admissions (n = 27,348) did not materially alter the primary finding: the age‐ and sex‐adjusted odds ratio for suboptimal image quality was 1.44 (95% CI 1.34–1.54; p < 0.001), compared with 1.47 (1.38–1.58) in the full cohort, and the age gradient was preserved (peak OR 1.57 at 65–74 years; attenuation to 1.19 [95% CI 0.98–1.44; p = 0.080] at ≥85 years). In an exploratory analysis of elective admissions (n = 4,503), where baseline mortality was lower (5.4%), the association was of larger magnitude (OR 2.25, 95% CI 1.72–2.95; p < 0.001), indicating that the prognostic signal is not driven by admission‐type heterogeneity (Table S14). Sex modified the association (interaction p = 0.028), stronger in males (IPTW OR 1.58) than females (1.35; Table S5). GBM models showed no incremental discrimination from NLP features (AUC 0.622 vs. 0.624). In the severity‐adjusted analysis (32,682 reports, 99.6%), SAPS‐II adjustment attenuated the overall OR from 1.50 to 1.35 (1.26–1.46) and OASIS to 1.31 (1.22–1.41), both highly significant (p < 10^{− 1 3}); the age‐stratified pattern was preserved, with effects elevated below age 85 and the ≥85 association attenuating to non‐significance (SAPS‐II‐adjusted OR 1.12, 0.91–1.37; Table S13).

Comorbidity burden explained only part of the age gradient. After adding the age‐independent Charlson comorbidity index to the age‐stratified models, the age‐by‐suboptimal interaction was attenuated and no longer statistically significant (p = 0.12), although stratum‐specific adjusted odds ratios retained the same overall shape (1.68, 1.55, 1.65, 1.42, and 1.24 from the youngest to the oldest stratum), and the overall suboptimal‐imaging effect was only marginally reduced (OR 1.53 to 1.50; comorbidity index OR 1.19 per point). Thus comorbidity accounts for a modest share of the age‐dependent pattern but does not abolish it, and the residual gradient coincides with a shift in goals of care. Recorded goals‐of‐care orders increased steeply with age: any code‐status limitation (DNR/DNI) rose from 7.9% of patients aged 18–49 to 32.0% of those aged ≥85, with comfort‐measures‐only orders following the same gradient (2.7% to 6.0%) (Figure 3). The age stratum in which the mortality association weakened was thus also the stratum in which goals‐of‐care limitations were most prevalent (Table 3).

FIGURE 3.

Goals‐of‐care orders and the comorbidity‐adjusted mortality effect across age strata. (A) Proportion of patients with any code‐status limitation (do‐not‐resuscitate/do‐not‐intubate, blue) and comfort‐measures‐only orders (red) by age group; both rise steeply with age, with code‐status limitations increasing roughly fourfold from 7.9% in patients aged 18–49 to 32.0% in those aged ≥85. (B) Odds ratios (95% CI) for the suboptimal‐imaging‐mortality association by age stratum after adjustment for the age‐independent Charlson comorbidity index; the effect is preserved through age 74 and attenuates in the oldest old, paralleling the goals‐of‐care gradient in panel A. The dashed line at OR = 1 indicates no association.

4. Discussion

In this large NLP‐based cohort, suboptimal echocardiographic image quality was independently associated with in‐hospital mortality, with comparable strength across all strata below age 85 and significant attenuation only in the oldest patients. Older patients also showed a qualitatively different reporting pattern even in technically adequate studies, driven by age at least as much as by image quality. The association persisted in a near‐complete severity‐adjusted sensitivity analysis (n = 32,682; 99.6%), and the ≥85 attenuation reinforces rather than undermines the geriatric framing of these results.

These findings are best understood through a geriatric lens. The clinical value of any diagnostic test is not fixed; it depends on what its result will change, and in older adults that value is shaped by life expectancy, functional status, and goals of care as much as by technical accuracy. A test that drives intervention in a fit 60‐year‐old may carry little prognostic consequence in a frail patient for whom invasive management is not intended. Our data make this measurable: the prognostic weight of suboptimal imaging was preserved wherever cardiac findings are likely to be acted upon (through age 74), softened in the transitional 75–84 stratum, and fell sharply in the oldest old. Critically, the attenuation in patients aged ≥85 is not evidence that image quality “matters less” biologically; it is what one would expect when competing mortality risks dominate and management is increasingly directed by goals of care rather than disease‐specific findings. Chronological age is, however, an imperfect proxy: two patients of the same age may differ profoundly in frailty, so the prognostic value of suboptimal imaging likely persists in fit individuals and disappears in frail ones. To probe this directly, we conducted two analyses linking the cohort to comorbidity and goals‐of‐care data. First, adjusting the age‐stratified models for the Charlson comorbidity index (age component removed) left the age‐by‐suboptimal interaction attenuated and no longer significant (interaction p = 0.12; stratum‐specific adjusted ORs 1.68, 1.55, 1.65, 1.42, and 1.24 from youngest to oldest), and the overall suboptimal‐imaging effect was only marginally attenuated (OR 1.53 to 1.50). The age gradient is therefore only partly explained by differential comorbidity burden, and a residual component remains. Second, recorded goals‐of‐care orders rose steeply with age: the proportion of patients with any code‐status limitation (DNR/DNI) increased from 7.9% in those aged 18–49 to 32.0% in those aged ≥85, a fourfold gradient mirrored by comfort‐measures‐only orders (2.7% to 6.0%) (Figure 3), supporting the interpretation that the attenuated association in the oldest old reflects a shift toward goals‐of‐care–directed management rather than a diminished biological link.

Prior work on suboptimal echocardiography has focused on prevalence, causes, and acoustic‐window improvement [4, 5, 27, 28]. To our knowledge, this is the first large‐cohort study to evaluate image quality as a prognostic variable rather than discarding suboptimal studies before analysis, with NLP enabling outcome analysis across the full adult age spectrum at a scale infeasible for manual review. The association persisted across unadjusted, adjusted, and propensity‐weighted analyses, and is best read as an epidemiological signal rather than a prediction tool: the modest discrimination (C‐statistic 0.61) means image quality flags a clinically informative marker rather than predicting mortality at the individual level.

The underlying mechanisms are likely multiple. Confounding by indication is the most probable: sicker patients are harder to image (hemodynamic instability, mechanical ventilation, positioning difficulty), and these same factors independently increase mortality—though our severity analysis bounded this contribution (10%–13% attenuation without abolition). The attenuated GEE estimate (OR 1.23 vs. IPTW 1.47) suggests further patient‐level confounding across repeated studies and represents a conservative lower bound. A direct causal contribution is also plausible, since incomplete assessment can leave important findings undetected (wall‐motion abnormalities, valvular pathology, pericardial effusion [29, 30]).

The age‐modified mortality impact is the most clinically relevant finding. No individual pairwise difference between adjacent strata survived multiple‐comparison correction; the categorical interaction (p = 0.04) therefore reflects a broad window of clinical vulnerability (IPTW ORs 1.54–1.66 below age 85) rather than a discrete age effect. Within this window echocardiographic findings drive high‐stakes decisions (severity grading of aortic stenosis and mitral regurgitation, ejection‐fraction categorization), so uncertainty from suboptimal imaging carries magnified consequences.

One apparent contradiction warrants clarification. The prevalence of “cannot be excluded” language rose monotonically with age (Table S7), yet in the multivariable model this construct was not independently associated with mortality (OR 0.982, p = 0.789; Table S12, Model C). The age‐related rise reflects a communication pattern shaped by image quality and complexity, whereas the multivariable null indicates that, once image quality is accounted for, the phrase adds no independent prognostic information—that is, diagnostic‐uncertainty language is downstream of image quality rather than an independent risk marker.

A secondary sex‐stratified analysis showed contrasting effects: women had more “cannot be excluded” language (OR 1.35, 1.23–1.49), whereas the mortality consequences of suboptimal imaging were stronger in men (IPTW OR 1.58 vs. 1.35; interaction p = 0.028; Table S5). Like the age interaction, this exploratory finding is hypothesis‐generating and warrants prospective study [31].

4.1. Robustness to Confounding

Illness severity, a known critical‐care confounder, was not directly captured in our primary adjustment set. We therefore ran a post‐hoc sensitivity analysis using two validated ICU severity scores (SAPS‐II [24] and OASIS [25]) from the MIMIC‐III concepts repository [16], available for 32,682 echocardiograms (99.6%). Suboptimal studies carried modestly higher baseline severity, confirming but bounding this confounding: after adjustment the overall odds ratio attenuated from 1.50 (95% CI 1.40–1.61) to 1.31–1.35 (Table S13) while remaining highly significant (p < 10^{− 1 3}). Severity therefore explains only a fraction of the association.

The age‐stratified pattern was preserved after severity adjustment, with effects elevated across all strata below age 85 (Table S13), while the ≥85 association attenuated to non‐significance (SAPS‐II‐adjusted OR 1.12, 0.91–1.37, p = 0.28). This supports rather than weakens our reading: in the oldest patients, mortality is dominated by underlying severity and competing causes of death, so the marginal contribution of image quality is correspondingly small.

The implied severity‐attributable confounding (RR ≈1.10–1.13) fell well below the E‐value of 2.02, supporting our primary benchmark. The ≥85 attenuation may also reflect the heterogeneity of the oldest‐old, in whom chronological age poorly surrogates biological vulnerability. A frailty‐stratified analysis (e.g., Clinical Frailty Scale), not feasible in MIMIC‐III, is needed to test whether the signal persists in fit ≥85 patients and attenuates in frail ones.

4.2. Clinical Implications

Treating image quality as a prognostic variable rather than an exclusion criterion is methodologically sound and aligned with calls for greater inclusion of older adults in cardiovascular evidence. Because the determinants of suboptimal imaging (body habitus, lung hyperinflation, thoracic geometry, tachycardia, hemodynamic instability) persist in contemporary practice, this reframing is conceptually robust to advances postdating MIMIC‐III, though the absolute effect magnitude should be treated as an upper bound pending replication in modern cohorts.

Our findings inform population‐level quality improvement rather than individual risk scoring, consistent with the modest discrimination (C‐statistic 0.61). Three age‐stratified priorities emerge. First, identify why a study is suboptimal, since the dominant cause is age‐dependent (correctable physiological factors in the young, fixed structural factors in the old) [4, 8, 32]. Second, calibrate the response to the age–risk gradient: in adults aged 18–74, where the association is consistently elevated (IPTW ORs 1.54–1.66), a low threshold for contrast or transesophageal echocardiography is warranted [9, 27, 33]; in the 75–84 group escalation should be individualized against multimorbidity; in those aged ≥85, emphasis shifts toward goals‐of‐care alignment, avoiding procedural burden unlikely to change management in frail individuals [9, 34].

Third, interpret the report by its age‐specific structure: the higher positivity ratio in older patients means a negative finding may not have been actively assessed, so clinicians should seek explicit confirmation when it is decision‐relevant. In this approach, image quality becomes a modifiable, age‐dependent determinant of diagnostic certainty rather than a fixed technical attribute.

4.3. Strengths and Limitations

Several design features strengthen these findings: a large cohort (32,813 reports, 19,731 patients) supporting fine‐grained age stratification with adequate power, a near‐perfect NLP classifier (κ = 0.996), and a robustness suite (IPTW with documented balance, GEE, E‐value sensitivity, formal interaction testing, and a near‐complete SAPS‐II/OASIS severity adjustment) addressing multiple sources of confounding and clustering.

This study has several limitations. Because MIMIC‐III derives from a single tertiary academic center with a predominantly critically ill population, reporting conventions and operator thresholds for flagging suboptimal image quality may differ from those at other institutions, and suboptimal imaging rates and their significance may differ in outpatient or community settings; external validation in a multicenter cohort with standardized image‐quality reporting is required before these findings can inform prognostic decision aids at other centers. The de‐identification protocol caps ages above 89 at 90, compressing the upper age distribution; a sensitivity analysis excluding these patients was consistent. Body mass index and standard severity scores (APACHE, SOFA) were unavailable; we addressed this through IPTW, GEE, E‐value sensitivity, and a post‐hoc SAPS‐II/OASIS‐adjusted analysis in 99.6% of the cohort that preserved the age‐stratified pattern. MIMIC‐III also lacks structured geriatric assessment variables (frailty, comprehensive geriatric assessment, polypharmacy, cognition); we used the Charlson index for disease burden and code‐status orders for goals‐of‐care intensity, but these do not capture frailty directly. Most fundamentally, this is an observational analysis and we cannot exclude that the association is wholly or partly explained by unmeasured confounding: suboptimal image quality may partly mark patient acuity that our adjustment set—despite IPTW, GEE, and severity scores—captures only incompletely, and the moderate E‐value (2.02) means a moderate‐strength unmeasured confounder could account for the estimate. We therefore frame image quality as a prognostic signal, not a causal factor. Our data could not link image quality to downstream clinical actions or post‐discharge outcomes. Report language is shaped by interpreter and institutional conventions. EchoGraph annotations were machine‐generated without human review, though the model's validation performance and recent peer‐reviewed publication [13] support confidence in aggregate statistics. Finally, MIMIC‐III predates several echocardiographic advances (harmonic imaging, contrast agents [27], 3D echocardiography), and external replication in a contemporary post‐2015 cohort is needed before findings can inform current guidelines. The interaction tests were pre‐specified but exploratory; the age interaction (p = 0.04) is of borderline significance and warrants confirmation.

5. Conclusions

Across 32,813 echocardiography reports from 19,731 patients, suboptimal image quality was common (27.9%), independently associated with in‐hospital mortality (IPTW OR 1.47), and not uniform across the age spectrum. Effect sizes were similar across early adulthood through the young‐old stratum (IPTW ORs 1.54–1.66, with overlapping confidence intervals), attenuating in the oldest patients as goals‐of‐care considerations shifted clinical priorities (IPTW OR 1.20, p = 0.003). A near‐complete sensitivity analysis adjusting for SAPS‐II and OASIS severity scores (n = 32,682; 99.6% of the cohort) preserved this age‐stratified pattern, with attenuation in the ≥85 group converging to non‐significance. For geriatric practice, these findings argue against discarding suboptimal studies as uninformative and instead support reading image quality as an age‐dependent marker of diagnostic certainty: one that warrants active escalation in adults under 85, but that should be weighed against frailty, competing mortality risks, and individualized goals of care in the oldest old. Pairing age stratification with formal frailty assessment in contemporary cohorts is the natural next step.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not‐for‐profit sectors.

Ethics Statement

This study used the de‐identified, publicly available MIMIC‐III database and the EchoGraph‐annotated ECHO‐NOTE2NUM derivative, accessed through PhysioNet credentialing. The Hacettepe University Health Sciences Research Ethics Board reviewed the protocol and determined that no formal ethics approval was required, as the data are anonymized and publicly available (no. SBA 26/199, June 24, 2026; ref. E‐16969557‐050.04‐00005034877).

Consent

The requirement for individual informed consent was waived. The study followed the Declaration of Helsinki.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Declaration of generative AI in scientific writing

The authors used a generative AI tool for language editing and take full responsibility for the content.

Supporting information

Table S1: Standard Logistic Regression vs. GEE Estimates.

Table S2: Standardized Mean Differences Before and After IPTW.

Table S3: E‐Values for Unmeasured Confounding.

Table S4: Temporal Sensitivity Analysis.

Table S5: Sex‐Stratified SuboptimalMortality Association.

Table S6: Comprehensive Sensitivity Analysis Summary.

Table S7: Diagnostic Uncertainty Language Prevalence by Age Group (%).

Table S8: “Cannot Be Excluded” and Mortality Stratified by Age and Image Quality.

Table S9: EchoGraph Entity Counts by Age Group (Mean±SD).

Table S10: NLP Classifier Validation Performance (n = 500).

Table S11: Suboptimal Quality by Age Group.

Table S12: Multivariable Mortality Regression Models.

Table S13: Severity‐Adjusted Sensitivity Analysis (SAPS‐II and OASIS).

Table S14: Admission‐TypeStratified Analysis of the Suboptimal ImagingMortality Association.

Figure S1: STROBE Patient Selection Flow Diagram.

Figure S2: Covariate Balance Before and After IPTW.

Figure S3: Calibration Plots for Mortality Models.

Figure S4: EchoGraph Entity Distribution by Age Group.

Figure S5: Comprehensive Sensitivity Analysis Forest Plot.

Data Availability Statement

The MIMIC‐III database is publicly available to credentialed researchers via PhysioNet (https://physionet.org/content/mimiciii/). The EchoGraph‐annotated ECHO‐NOTE2NUM dataset is available from its original source as cited. Analysis code is available from the corresponding author on reasonable request.