Spinal muscular atrophy (SMA) is an autosomal recessive disorder caused by biallelic loss of the survival motor neuron 1 (SMN1) gene [1]. Since 2016, three disease-modifying therapies have been approved, all now available to adults. The antisense oligonucleotide nusinersen and the oral splicing modifier risdiplam benefit adults as well as children [2]. The gene-replacement therapy onasemnogene abeparvovec, previously limited to infants, reached patients aged 2 and older through an intrathecal formulation (2025). Adults now make up roughly half the prevalent population, an estimated 8000‐10,000 individuals [3]. Adults first diagnosed in childhood, when genetic confirmation was not routine and before treatment existed, may have had little reason to follow up, especially if mild or stable. Identifying them within a health system now relies on coded electronic health record (EHR) data.

The primary SMA diagnostic code is an unreliable marker. Diagnostic terminologies lack SMA-specific codes and group distinct disorders under shared rubrics, conflating true SMA with similarly named conditions such as spinal and bulbar muscular atrophy (SBMA) [4]. In a claims-based cohort, SMA codes did not distinguish subtypes or confirm true cases [5], and in a genomic cohort, overlapping codes could not confirm SMN1-SMA without genetic testing [6].

We quantified the positive predictive value (PPV) of SMA-associated diagnostic codes against chart adjudication in one US health system and characterized the miscoded cases.

Of 60 charts, 22 (37%, 95% CI 26%‐49%) were appropriately coded for SMA or a related motor neuron disorder. Sixteen (27%, 95% CI 17%‐39%) had molecularly confirmed SMA, and the other 6 were appropriately coded motor neuron conditions (5 Hirayama disease, 1 distal SMA). The PPV of an SMA-associated code for molecularly confirmed SMA was therefore 27%, rising to 37% if related motor neuron disorders are included. Among the 16 confirmed patients, all seen after 2016, 10 (63%) had no documented disease-modifying therapy.

The remaining 38 charts (63%) were miscoded (Table 1), comprising conditions unrelated to SMA (n=10, 17%), asymptomatic SMA carriers (n=9, 15%), charts with no relevant neuromuscular diagnosis (n=9, 15%), SBMA (n=5, 8%), and prenatal carrier screening (n=5, 8%).

Most adults whose chart carried an SMA code did not have SMA. Fewer than 1 in 3 had confirmed disease, well below the roughly 4 in 5 correct for the infantile SMA code in a national registry [7], the two-thirds for amyotrophic lateral sclerosis codes [8], and the one-half for muscular dystrophy [9]. No comparable estimate existed for the SMA-associated codes used in adults. Because the the ICD-10 codes provide no SMA subtype and bundle genetically distinct disorders, carriers, and screening encounters under shared rubrics, the code cannot by itself separate affected patients from unaffected ones.

Although SBMA and distal SMA can resemble SMA clinically, they are distinct genetic diseases. SBMA lacks a dedicated code and is often filed under SMA or adjacent codes, defensible administratively yet reflecting a classification ICD-10 no longer captures. Asymptomatic carriers and prenatal screening encounters represent unaffected individuals. Counting any as SMA inflates the apparent cohort; here, miscoded charts outnumbered confirmed cases more than 2 to 1, so a registry built on these codes would be wrong before its first analysis. Because coding underlies disease queries, this undermines estimates of prevalence, treatment patterns, and natural history from single systems to national registries.

Correcting this requires two changes. First, the codes must be aligned with the clinical record so they are applied as documented and anchored on molecular SMN1 confirmation with type-specific subcodes so they are precise. Europe built Orphacodes, a dedicated rare-disease coding system, to capture diseases ICD-10 cannot distinguish [10], and a comparable US reform through the National Center for Health Statistics could require SMN1 confirmation for the main SMA code, add subcodes for types 0 to 4, and retire the catchall categories. Second, these codes must then be applied consistently to every note, old and new. In future work, disease-specific natural language processing could recode existing records and check new entries against the documented diagnosis.

As a single-center, retrospective estimate from a small sample, this rate may not transfer to other settings, though the pattern appears in claims data elsewhere. Adjudication drew on records in which every patient had been evaluated by at least two neurologists, and confirmed cases were anchored on molecular testing rather than reviewer judgment. Because all confirmed patients were seen after 2016, undocumented therapy is not a pretreatment artifact of limited confirmation or stable disease; it more likely reflects care outside this system, which a single institution cannot distinguish from loss to follow-up. Privacy-preserving record linkage could close this gap.

Despite these limits, SMA-associated diagnostic codes are an unreliable marker for confirmed disease. Codes that are aligned, molecularly anchored, consistently applied, and audited would make identification reliable, strengthening prevalence estimates, treatment tracking, and natural-history research.