Innovation trajectory of Software as a Medical Device: Evidence from the US FDA-approved products

Abstract

Background

Digital health innovation is reshaping healthcare, with Software as a Medical Device (SaMD) as a critical frontier. However, little is known about how enterprises evolve portfolios within this regulated market.

Objectives

This study uses enterprise-level SaMD product portfolio data to characterize innovation trajectories in SaMD development and assess heterogeneity in these trajectories across enterprise characteristics.

Methods

We identify 578 FDA-approved products from 289 enterprises (2012–2022) using the OpenFDA database. A multi-state model tracked innovation dynamics across 4 states: (1) one initial approval, (2) multiple approvals in identical area, (3) expansion into 2 areas, and (4) diversification into 3 or more areas. We apply Kaplan–Meier survival analysis with log-rank tests, Competing-risk cumulative incidence analysis, and Cox proportional hazards regression to examine how enterprise size and industry classification influenced portfolio expansion, with right-censoring at the observation end date.

Results

The analysis reveals stagnating innovation trajectories. 67% (n=193) of enterprises made only a single FDA submission. Kaplan-Meier estimation yielded a median time to first portfolio expansion of 8.41 years over a 10-year observation window, with large enterprises transitioning significantly faster than micro enterprises ( $p$ <0.001). Neither industry nor entry timing showed significant differences. Cumulative incidence functions show that the five-year cumulative incidence of focused expansion (33.6%) was five times that of diversification (6.4%). Cox proportional hazards regression confirms enterprise size as the dominant predictor (micro vs. large HR=0.19, $p$ <0.005). Software enterprises show a significantly higher expansion hazard (HR=2.15, $p$ =0.04), while other industry types showed no significant effect.

Conclusions

Innovation of SaMD follows structurally constrained trajectories. We propose that this is caused by resource barrier that hinders smaller enterprises. Overcoming this barrier requires innovation systems that facilitate integration of complementary assets. Sustainable innovation depends on shifting from isolated product development to coordinated, multi-domain ecosystems.

Introduction

The healthcare industry is undergoing a profound digital transformation (DT) driven by technologies like artificial intelligence (AI) and the Internet of Things (IoT). Similar to digital innovation in other sectors, healthcare DT leverages the unique affordances of digital technology, like reprogrammability, interoperability, and generativity, to enable continuous recombination and adaptation across diverse clinical contexts.^1,2 This transformation marks a paradigmatic shift in the way healthcare services are designed and delivered. What was once a conventional, clinician-led process has evolved into a digitally mediated system in which patient behaviors, data analytics, and software technologies interact dynamically to redefine medical practice, organizational processes, and value creation mechanisms.^3–5

At the center of this transformation lies SaMD, a representative form of digital innovation that encapsulates the convergence of algorithms, data, and clinical decision-making.^6,7 Defined as standalone software that performs medical functions without being part of a hardware device, SaMD encompasses applications ranging from diagnostic algorithms to treatment planning and remote monitoring systems.^8,9 Its digital nature allows for rapid iteration, scalability, and seamless integration into existing health information infrastructures, thereby enhancing responsiveness to emerging healthcare needs.^10,11 Furthermore, SaMD unlocks new pathways for innovation and value creation by enabling predictive diagnostics, continuous monitoring, and personalized treatment recommendations that transcend the limitations of traditional medical devices.^6,12,13 Despite the growing significance of SaMD, systematic understanding of how enterprises build and develop their product portfolios under regulatory uncertainty remains limited. This study addresses this gap by examining enterprise-level innovation trajectories, investigating the heterogeneity of enterprises’ innovation trajectories.

The extant literature on SaMD has predominantly clustered around two domains: addressing the technical challenges in delivery and revealing how the existing regulatory framework should adapt to technological changes. Studies on regulatory frameworks, particularly those addressing AI-based medical software, have examined critical issues like algorithmic transparency, post-market surveillance requirements, and the adequacy of existing regulatory pathways for adaptive algorithms.^10–12,14 A parallel stream of research has focused on clinical validation, investigating efficacy, safety, and usability concerns, thereby emphasizing the technical and medical dimensions of SaMD deployment.^12,15–17 While these studies provide valuable insights into compliance and clinical performance, relatively few have adopted an innovation management perspective to understand how enterprises strategically navigate innovation in regulated healthcare markets. This perspective is particularly critical for SaMD, as it occupies a unique tension between the rapid, iterative nature of digital development and the rigid, high-stakes requirements of regulation. Unlike conventional hardware-based medical devices, SaMD requires continuous updates and real-world data integration, yet these must occur within a framework designed for stability and risk aversion. For instance, the FDA’s real-world evidence framework requires developers to demonstrate clinical validity through systematic collection and analysis of real-world data for each new application, while the EU MDR (2017/745) classifies most SaMD as Class IIa or higher, mandating Notified Body involvement and treating each algorithm change as a major modification requiring re-evaluation. These cumulative documentation and compliance requirements multiply with each new application area, creating escalating barriers to portfolio diversification that fall disproportionately on smaller enterprises. There remains a lack of systematic analysis of enterprise-level factors that shape innovation in SaMD. This paucity of enterprise-centric evidence limits our understanding of how innovation unfolds under regulatory constraints and leaves practitioners without empirically grounded frameworks for managing SaMD product development.

This study traces trajectory dynamics of enterprise by applying the concept of innovation trajectory. ¹⁸ Trajectory describes the path or progression of states, behaviors, or actions (hereafter outcome) over time. ¹⁹ The pioneering studies by Dosi (1982) and Abernathy & Utterback (1978) define technological trajectory which show that technological development follows distinct paths over time and shapes competitiveness.^20,21 Follow this logic, Juliao-Rossi et al. (2020) define the innovation trajectory, in the context of innovation, as the trajectory of innovation is a set of states of innovation followed by a firm over time and extend this concept to the firm level, documenting heterogeneous longitudinal pathways in firms’ innovation output shaped by organizational capabilities and structural constraints. This perspective is well suited to digital health and SaMD, where innovation evolves alongside regulatory change. This study identifies the trajectory by tracing these portfolio-level dynamics to capture how firms sequentially produce innovations under external constraints. We aim to build theoretical understanding of how innovation in SaMD unfolds as a trajectory jointly shaped by enterprise behavior and regulatory co-evolution. This provides the analytical foundation for identifying structural barriers and for formulating strategies and policies that foster sustainable digital innovation in healthcare. Accordingly, this study addresses 2 research questions (RQs):

RQ1: What innovation trajectories characterize enterprise-level SaMD product portfolio development?

RQ2: How does heterogeneity in enterprise characteristics shape SaMD innovation trajectories?

Method

Data collection

This study is a retrospective cohort study of enterprises whose SaMD products are approved by the U.S. Food and Drug Administration (FDA), covering the period January 2012 to February 2022. U.S. FDA is the sole regulatory authority for commercially marketed medical devices in the United States. The FDA’s OpenFDA database serves as the primary source of information, providing comprehensive details on approved medical devices, including device descriptions, indications for use, and decision summaries.^22–24

To identify SaMD products and corresponding enterprises, we followed the data collection method used in the previous studies. ²⁵ This process involved downloading product information for 44,846 FDA-approved medical devices in PDF and HTML formats, covering the period from January 2012 to February 2022. The selection of January 2012 as the starting point aligns with the International Medical Device Regulators Forum’s (IMDRF) initial discussions on the SaMD definition. ⁹ We employed a keyword-based filtering approach to identify potential SaMD products. The keywords utilized included ‘Software as Medical Device’, ‘Software as a Medical Device’, ‘Standalone software’, ‘Software only’, ‘Software package’, and ‘Software device’. This broad keyword set was designed to prioritize recall and minimize false negatives, as SaMD products typically highlight these technical characteristics in their FDA approval applications. This initial screening yielded 723 devices, a figure intended to be over-inclusive. Subsequently, 2 authors conducted manual filtering. One author pre-scanned the documentation of the 723 devices to calibrate the understanding of diverse product descriptions. Then the author manually checked if the documentation claims itself as SaMD following a simple criterion that a device was retained only if its official FDA regulatory documentation formally claimed SaMD status. Devices that merely mentioned SaMD in a general product description were excluded. The other author follows the same criterion and ran a double check to confirm the filtered devices. Finally, we identified 581 devices eligible for classification as SaMD, effectively filtering out false positives like software components integrated into hardware.

Given the absence of a comprehensive database containing profiles of enterprises, we conducted manual searches across multiple reputable platforms, including Crunchbase, Bloomberg, PitchBook, and individual enterprise websites. This meticulous process allowed us to compile comprehensive enterprise profiles, encompassing details like establishment date, focus industry, and employee count. Crunchbase is recognized as a comprehensive database for startups, ²⁶ Bloomberg, a global financial information provider, ²⁷ and PitchBook, ²⁸ specializing in private market data, served as primary sources for this information. Crunchbase served as the primary data source due to its specialized focus on startup and technology enterprise profiles. In cases where data were missing or required further verification, Bloomberg and PitchBook were utilized as supplementary sources. To ensure data accuracy, we cross-referenced information across these platforms. In the event of minor discrepancies in dynamic data (e.g., employee counts), we prioritized the most recent updates from Crunchbase or verified the information against the enterprise’s official annual reports and websites. In alignment with a focus on enterprise behavior, this study conducted manual verification of the FDA approval submitters’ profiles. This additional step led to the exclusion of 2 entities identified as government and university institutions, along with their corresponding 3 SaMD. Consequently, the final dataset used for analysis consists of 578 SaMD products developed by 289 enterprises, ensuring our study accurately reflects the landscape of commercial SaMD development and innovation (Figure 1).

Figure 1.

Data extraction and aggregation approach.

Data preprocessing

To examine how enterprise characteristics shape SaMD innovation trajectories, we first standardized enterprise-level attributes used throughout the survival analysis. We first note that there is no globally accepted standard for classifying enterprise size. We therefore adopted a four-tier classification following the OECD four-tier definition. ²⁹ The. The resulting categories are: micro (1-10), small (11–50), medium (51–250), and large (≥251). The micro category is distinguished from small to capture the distinct dynamics of early-stage enterprises and startups. Because the data sources report employee counts in predefined ranges (e.g., 1–10, 11–50, 51–250) rather than exact headcounts, the classification boundaries differ slightly from the standard OECD cutoffs (<10, 10–49, 50–249, ≥250) (Appendix Table 1). We determined industry classification using each enterprise’s primary business sector as categorized following the Bloomberg Industry Classification Standard, encompassing sectors like Medical Equipment & Devices, Software, Technology Services, and Biotech & Pharma. We operationalized portfolio diversification as the number of distinct FDA application areas in which an enterprise had obtained SaMD approvals, serving as the basis for state assignment in the multi-state model. For each enterprise, we tracked cumulative counts of submissions, distinct medical subjects, distinct regulatory regimes, and distinct application areas at each submission event.

We conducted a retrospective cohort study of 289 enterprises developing SaMD, using FDA approval records from January 2012 to February 2022. For enterprises founded after January 2012,follow-up began at the date of foundation; all enterprises were right-censored at the end of the observation period in February 2022. We defined a multi-state model with four states representing progressive levels of product portfolio diversification across medical application areas (Table 1). An enterprise entering State 1 with its first SaMD approval could subsequently expand its portfolio to obtain multiple approvals in the identical application area (State 2), or in two different application areas (State 3), or remain in State 1 with only one FDA approval (State 1). Enterprises that continued to diversify could gain approvals across three or more different application areas (State 4). This framework allows transitions from lower to higher states, including direct transitions that bypass intermediate states, and captures both sustained innovation within existing application areas and diversification into new medical domains. To support the survival analysis, we restructured the event-level submission data into an episode-level dataset, where each record represents a company’s sojourn in a given state. We treated repeated submissions within the same state (e.g., multiple approvals in a single application area while remaining in State 2) as continued sojourn rather than state transitions; only changes in the number of distinct application areas triggered a state transition. We defined entry time as the date of arrival in that state and exit time as the date of the subsequent state transition or, for enterprises that did not transition, the end of the observation period. For each episode, we recorded a binary transition indicator in a right-censored way which is set as 1 if the enterprise moved to a higher state, 0 if it remained in the current state until the end of observation. In the survival analysis framework explained in following section, we refer observed transitions as events. This restructuring yielded 393 episodes across 289 enterprises. For enterprises with only a single FDA submission (n=193), we assigned a censoring time equal to the interval between their submission date and the observation end date, rather than a duration of zero, so that right-censoring was handled correctly in the survival models. We verified temporal ordering and state-logic conformity for all enterprises through manual data checks.

Table 1.

State definition of enterprises in SaMD development.

State	Definition
1	Achieved the first SaMD approval in a single application area
2	Achieved multiple SaMD approvals in the identical application area as the first approvals
3	Achieved multiple SaMD approvals in 2 different application areas
4	Achieved multiple SaMD approvals in 3 or more different application areas

Multi-state survival analysis

We implemented all analyses in Python 3.13 using the lifelines (v0.30) and scipy (v1.16) libraries. We adopted a multi-state survival framework to capture the timing, direction, and competing nature of post-entry portfolio expansion, dimensions that a single-endpoint regression collapses into a binary outcome. Within this framework, we applied three complementary approaches, each addressing a distinct question. Kaplan–Meier estimation answers when enterprises expand after first approval. Cumulative incidence functions (CIF) answer which competing pathway they follow, either focused repeat (State 2) or diversification (State 3), under proper competing-risk adjustment. Cox proportional hazards (Cox PH) regression identifies which enterprise- and industry-level covariates drive the hazard of expansion. First, we estimated transition-specific survival functions using the Kaplan–Meier method to characterize the time from first FDA submission to subsequent portfolio expansion. We conducted stratified analyses by enterprise size, industry classification, and entry cohort (2012–2016 vs. 2017–2022), with pairwise log-rank tests and Bonferroni correction for multiple comparisons. Second, we estimated CIF to account for the competing risk structure. Because enterprises exiting State 1 face two competing destinations, State 2, focused expansion within the same application area, or State 3, diversification into a new area. Unlike standard Kaplan–Meier estimation, which overestimates the probability of each event type when competing risks are present, the CIF partitions the overall transition probability among competing destinations by incorporating the cause-specific hazard at each event time weighted by the overall survival probability. We used bootstrap resampling with 1,000 iterations to construct 95% confidence intervals (CI) for the CIF estimates and performed stratified CIF analyses by enterprise size and industry to compare pathway probabilities across subgroups. Third, we fitted Cox PH regression to identify predictors of portfolio expansion while controlling for potential confounders. Time-fixed covariates included enterprise size (micro, small, medium with large as reference), industry type (software vs. non-software, medical device vs non-medical device), company age at market entry (continuous, in years), and entry cohort (2012–2016 vs. 2017–2022). For industry type, Technology Services, Biotech and Pharma, Health Care Facilities and Services, Technology Hardware, and Commercial Support Services were combined as reference due to insufficient event counts for stable estimation industry type. We encoded enterprise size as three indicator variables (medium, small, micro) relative to the large reference category. We assessed the proportional hazards assumption using Schoenfeld residual tests for each covariate and limited the number of covariates following the ten events-per-variable rule. We also fitted cause-specific Cox PH regression separately for each competing destination from State 1 to isolate the covariate effects on focused versus diversified expansion pathways. We reported later-stage transitions descriptively owing to insufficient event counts for formal inference.

Results

Descriptive statistics

Between January 2012 and February 2022, a total of 289 enterprises obtained 578 FDA approvals (Appendix Figure 8), with large enterprises securing the highest number of approvals per enterprise (Table 2; Appendix Figure 7).

Table 2.

Distribution of enterprises who get US FDA approved SaMD products by category.

Total	289
By Enterprise size of Applicant Enterprise
Micro	55 (19.0%)
Small	113 (39.1%)
Medium	61 (21.1%)
Large	60 (20.8%)
By Main Industry of Applicant Enterprise
Medical Equipment & Devices	119 (41.2%)
Software	128 (44.3%)
Technology Services	15 (5.2%)
Biotech & Pharma	11 (3.8%)
Health Care Facilities & Services	9 (3.1%)
Technology Hardware	4 (1.4%)
Commercial Support Services	3 (1.0%)
By State of Applicant Enterprise
First SaMD submission (1 application area)	193 (66.8%)
Additional SaMD submission (1 application area)	76 (26.3%)
Additional SaMD submission (2 application areas)	17 (5.9%)
Additional SaMD submission (>= 3 application areas)	3(1.0%)

The data in brackets are percentages (%) for categorical variables.

Descriptive breakdown of 289 enterprises by size (micro, small, medium, large), main industry, and final portfolio state at the end of the observation period. Includes enterprise counts and percentages for each category.

In terms of enterprise size, large enterprises (n=60) account for 41.9% of approvals (n=242), followed by medium enterprises (n=61), which account for 17.6% of approvals (n=102), small enterprises (n=113) account for 29.6% (n=171), and micro enterprises (n=55) account for 10.9% of approvals (n=63). Overall, the distribution of FDA approvals by enterprises was highly skewed. 66.8% (n=193) of enterprises hold only 1 FDA approval, whereas only 3.8% (n=11) of enterprises have more than 5 FDA approvals (Appendix Figure 9). Despite micro and small enterprises collectively contributing 40.5% (n=234) of total approvals, each of these enterprises holds an average of 1.4 FDA approvals, significantly lower than large enterprises (4.0 approvals per enterprise) and medium enterprises (1.7 approvals per enterprise). Siemens, GE Healthcare, Varian, Ewoosoft, and Philips, the top five large enterprises, account for 47.1% (n=114) of approvals by large enterprises.

In terms of industry classification, software enterprises lead (n=128) account for 38.8% approvals (n=224), followed by medical equipment and device manufacturers (n=119) account for 50.7% approvals (n=293), technology service providers (n=15) account for 4.3% approvals (n=25), and other industries (n=27) account for 6.2% approvals (n=36). (Appendix Table 2, Appendix Table 3).

In terms of product portfolio, enterprises intend to concentrate on specific application areas and medical specialties. Most enterprises (93.1%, n=269) specialized in a single application area, while 5.9% (n=17) received approvals in two areas. Only 1% (n=3) obtained approvals across more than two application areas. A similar pattern was observed in medical specialties, where 96.9% (n=280) of enterprises had approvals in only one specialty, and just 3.1% (n=9) covered two specialties. Examining diversification across industry classifications, we found that while most enterprises had focused on a single application area, those in Medical Equipment & Devices and Software had demonstrated a higher tendency for diversification. Among Software enterprises (n=128), 9 enterprises covered two or more application areas, including 2 large enterprises, 4 medium enterprises, 2 small enterprises, and 1 micro enterprise. In Medical Equipment & Devices (n=119), 10 enterprises had approvals in two or more application areas, with 9 being large enterprises and 1 medium enterprise. However, only 3 enterprises, General Electric (GE), Siemens AG, and Koninklijke Philips N.V., expanded beyond 3 application areas. Regarding medical specialty coverage, diversification remained limited. Only 7 large enterprises in Medical Equipment & Devices and 2 large enterprises in the Software industry obtained approvals in 2 medical specialties. No enterprise covered more than 2 medical specialties.

SaMD innovation trajectory patterns

The multi-state transition diagram (Figure 2) maps the innovation trajectories of all 289 enterprises. Of the 289 enterprises entering State 1, 81 (28%) transitioned to State 2 by obtaining additional approvals in the same application area, while 15 (5%) moved directly to State 3 by expanding into a second application area. The remaining 193 (67%) made no further submissions. Among the 81 enterprises reaching State 2, only 5 (6%) subsequently diversified to State 3, and of the 20 enterprises in State 3, only 3 (15%) progressed to State 4.

Figure 2.

Multi-state transition diagram of SaMD enterprise portfolio evolution.

Most enterprises remained in their initial state for an extended period following their first FDA submission. The Kaplan-Meier estimated median time to first transition was 8.41 years (95% CI: 6.17 years to not reached), and approximately 60% (95% CI: 53%∼66%) of enterprises had not made a second submission after five years (Figure 3). All transition rates in this section are Kaplan–Meier estimates adjusted for right-censoring unless stated as observed counts. Of the 289 enterprises in the cohort, 193 (66.8%) never made a second submission during the observation period. The at-risk population fell from 289 at baseline to 160 at two years and 85 at four years; the first two years after market entry thus concentrated the majority of transition activity. After this window, transitions were sparse, with only 6 enterprises remaining at risk by the eighth year. Estimates beyond five years should be interpreted with caution given the small at-risk population.

Figure 3.

Kaplan–Meier survival curve for time to first portfolio expansion.

When stratified by enterprise size, expansion rates diverged sharply (Figure 4(a)). Kaplan-Meier estimated five-year transition rates were 56% (95% CI: 43%∼70%) for large enterprises (n=60; 30 observed transitions), 37% (95% CI: 27%∼49%) for small (n=113; 35 observed), and 17% (95% CI: 7%∼35%) for micro (n=55; 6 observed). CI of micro is wide given the small number of events. At two years, 22 of 60 (37%) large enterprises had already transitioned, compared to 4 of 55 (7%) micro enterprises. Pairwise log-rank tests were significant for large vs. micro ( $p$ <0.001), medium vs. micro ( $p$ =0.005), both below the Bonferroni-corrected threshold of 0.0083. The micro vs. small ( $p$ =0.013) and large vs. small ( $p$ =0.026) did not survive correction.

Figure 4.

Kaplan–Meier survival curves stratified by enterprise characteristics. (a) Stratified by enterprise size (large, medium, small, micro). Curves diverge sharply, with large enterprises transitioning fastest. (b) Stratified by industry (software, medical equipment and devices, other). No significant differences observed. (c) Stratified by entry cohort (2012–2016 vs. 2017–2022). No significant differences observed. Each panel includes at-risk tables.

Stratification by industry (Figure 4(b)) produced no significant differences among software enterprises (n=128), medical equipment and device manufacturers (n=119), and other industries (n=42), with nearly identical survival curves (all pairwise $p$ >0.05). Stratification by entry cohort (Figure 4(c)) similarly showed no difference between enterprises (n=105) entering in 2012–2016 and those (n=184) entering in 2017–2022 ( $p$ =0.58).

Cox PH regression, controlling for industry, company age at entry, and entry cohort, identified enterprise size as the dominant predictor (Appendix Figure 1; log-likelihood ratio test $p$ = 0.0015, concordance=0.64). One enterprise was excluded from the regression due to missing foundation year data, yielding an analytical sample of 288. The log-likelihood ratio test evaluates whether the model with all covariates fits significantly better than a null model; the concordance index measures discriminative ability (0.5 = no discrimination, 1.0 = perfect), with 0.64 indicating moderate discrimination consistent with the inherent heterogeneity of enterprise innovation behavior. Micro enterprises had a hazard ratio of 0.19 relative to large enterprises (95% CI: 0.08–0.48, $p$ <0.005), meaning their expansion rate was roughly one-fifth that of large firms. Small enterprises had a hazard ratio of 0.49 (95% CI: 0.29–0.85, $p$ =0.01), and medium enterprises showed a non-significant reduction (HR=0.64, 95% CI: 0.37–1.11, $p$ =0.11). Software enterprises showed a significantly higher hazard of expansion (HR=2.15, $p$ =0.04). In cause-specific models, the software effect was not significant for focused expansion (State 1 to 2, HR=1.58, $p$ =0.22), suggesting it is driven by the diversification pathway. Medical equipment and devices (HR=1.54, $p$ =0.23), company age (HR=1.00, $p$ =0.90), and entry cohort (HR=1.00, $p$ =0.99) did not reach significance, ruling out confounding by these factors. Schoenfeld residual tests supported the proportional hazards assumption for all size covariates (all $p$ >0.10); only company age showed a violation ( $p$ =0.003) (Table 3). Because company age at entry was not a primary predictor (HR=1.00, p=0.90), this violation does not affect the substantive conclusions. Even when restricted to the 96 enterprises with two or more submissions, large enterprises reached State 3 or higher at a rate of 36.7% (n=11), versus 20.0% (n=5) for medium, 16.7% (n=1) for micro, and 8.6% (n=3) for small enterprises (Figure 5). This pattern points to a structural constraint tied to organizational scale rather than a deliberate strategic choice.

Table 3.

Cox proportional hazards regression for portfolio expansion from State 1.

Covariate	Any transition (n=288, 96 events)			Cause-specific: State 1 to 2 (n=288, 81 events)
	HR	95% CI	$p$	HR	95% CI	$p$
Enterprise size (ref: Large)
Medium	0.64	0.37-1.11	0.11	0.73	0.39-1.34	0.31
Small	0.49	0.29-0.85	0.01	0.62	0.34-1.12	0.11
Micro	0.19	0.08-0.48	<0.005	0.21	0.08-0.56	<0.005
Industry type (ref: Other)
Software	2.15	1.04-4.46	0.04	1.58	0.76-3.29	0.22
Medical equipment & devices	1.54	0.76-3.15	0.23	1.20	0.58-2.49	0.62
Company age at entry (continuous, years)	1.00	0.99-1.01	0.90	1.00	0.99-1.01	0.68
Entry cohort (ref: Early entry 2012-2016)
Late entry 2017-2022	1.00	0.65-1.54	0.99	1.08	0.67-1.72	0.76

Hazard ratios, 95% confidence intervals, and p-values for six covariates across two models: any first transition from State 1 (96 events) and cause-specific transition to State 2 (81 events). Reference category for enterprise size is large. The cause-specific model for State 1 to State 3 (15 events) is omitted due to limited statistical power. One enterprise was excluded due to missing foundation year (company age covariate unavailable), reducing the analytical sample from 289 to 288.

Figure 5.

Enterprise size versus final portfolio state.

Enterprise heterogeneity and trajectory divergence

Enterprises leaving State 1 faced two competing destinations, and cumulative incidence analysis showed a clear asymmetry between them (Appendix Figure 2(a)). Of 289 enterprises, 81 moved to State 2, focused expansion in the same application area, 15 moved to State 3, diversification into a new area, and 193 remained in State 1. The five-year cumulative incidence of focused expansion was 33.6% (95% CI: 27.3%–40.4%); that of diversified expansion was 6.4% (95% CI: 3.3%–9.8%). The CI did not overlap, placing focused expansion at roughly five times the probability of diversification. Focused expansion accumulated quickly in the first two years, while diversified expansion accrued more evenly across the observation period. When stratified by enterprise size (Appendix Figure 2(b)), large enterprises had higher cumulative incidence on both pathways: approximately 60% for focused expansion at five years, versus roughly 25% for medium and small enterprises and under 10% for micro enterprises. The diversified expansion CIF stayed below 5% for all categories except large. Industry stratification (Appendix Figure 2(c)) showed no meaningful differences; software, medical equipment, and other industries had overlapping curves. Organizational scale thus shaped both the likelihood and the direction of expansion, while industry type did not.

Transitions beyond the initial pathway were rare. Of the 81 enterprises that reached State 2 through focused expansion, only 5 (6.2%) later moved to State 3 by entering a second application area; the median time spent in State 2 before this transition was 559 days (Appendix Figure 3). Three of the five were large enterprises. Of the 20 enterprises in State 3, only 3 (15%) progressed to State 4, with a median sojourn of 456 days. All three (GE Healthcare, Philips, and Siemens) were large enterprises with more than 10 FDA submissions across three or more application areas. Broad portfolio coverage, in practice, was limited to these organizations. Given the small number of events, these later-stage transition estimates carry wide CI and should be regarded as descriptive rather than inferential. Most enterprises, once they reached a given state, remained there for the rest of the observation period.

At the ecosystem level, the state composition of the SaMD market changed little over the decade, even as the market itself grew rapidly (Figure 6). The number of active enterprises grew from a single enterprise in 2012 to 289 by 2022, but this growth came almost entirely from new State 1 entrants. The share of enterprises in State 1 fell only from about 75% to 67%; the State 2 share grew from 15% to 26%; States 3 and 4 combined by the end of observation did not exceed 10%. The market expanded by adding single-product entrants, not by existing participants maturing and diversifying. The Sankey diagram (Appendix Figure 6) traces this pattern across milestones: at one year after entry, 255 of 289 enterprises remained in State 1; by the final observation, 193 still had not transitioned.

Figure 6.

State composition of the SaMD enterprise ecosystem over time.

Visualization of all 289 enterprise trajectories over time confirmed that stagnation was the dominant pattern (Appendix Figure 4, Appendix Figure 6). In the trajectory heatmap, most enterprises held a single state from entry to the end of observation, with only a narrow band at the top (those reaching States 3 and 4) showing progression across multiple states. Among the 20 most active enterprises by submission count (Appendix Figure 5), trajectories varied widely. Siemens, the most active with 48 approvals, moved through all four states over nearly nine years. GE Healthcare (23 approvals) reached State 4 within two years of its first submission. Varian, by contrast, filed 16 submissions yet never left State 2, concentrating on a single application area. We treat Varian as an independent enterprise because the majority of its submissions predate its April 2021 acquisition by Siemens Healthineers. Philips (12 approvals) similarly reached State 4 within its first year, sustaining portfolio breadth over a five-year span. These cases illustrate that submission volume alone does not produce diversification: Varian’s 16 submissions did not achieve the breadth of GE Healthcare’s 23. Diversification appears to require not just sustained activity, but a breadth of organizational capability concentrated among the largest firms.

Discussion

Stagnating innovation trajectory of SaMD

The transition probabilities shown in Figure 2 suggest that the innovation trajectory of SaMD enterprises is stagnating and lacks consistent upward progression from single to diversified and serial offerings. We identified 3 trajectories of digital innovation in healthcare (Table 4). 193 enterprises only have 1 FDA-approved SaMD. 76 enterprises have multiple FDA-approved SaMD in an identical application area. Merely 20 enterprises diversified and had multiple FDA-approved SaMD in multiple application areas. Instead of following a sequential pathway of expanding into multiple application areas, enterprises were frequently consolidated within their initial domains before taking measured steps toward diversification. Enterprises exhibited trajectory persistence shaped by structural constraints, where resource limitations rather than strategic choice largely determined whether further diversification occurred.

Table 4.

Trajectory patterns of digital innovation in healthcare.

Trajectory pattern	Description	Number of enterprises	Typical characteristics
Stay-to-initial	Enterprises remain in State 1 with a single SaMD approval in one application area	193	Predominantly, startups and new entrants with limited access to specialized resources
Focused-Expansion	Enterprises achieve State 2 with multiple SaMD approvals in an identical application area	76	A mix of startups and established players with sufficient resources for depth but not breadth
Diversified-Expansion	Enterprises reach States 3 and 4 with SaMD approvals across multiple application areas	20	Predominantly incumbents with established access to diverse healthcare resources

The state transition diagram (Figure 2) provides empirical insight that most enterprises remain within their established domains, with only a subset successfully transitioning to broader portfolios through incremental progression. For example, fewer transitions from State 1 to State 3 indicate that most enterprises followed incremental transitions. Additionally, the relatively high probability of remaining in State 2 suggests that enterprises have not progressed beyond their initial focus areas, reinforcing the pattern of specialization rather than broad portfolio expansion. Although this study focuses on products that successfully obtained FDA approval, a clear pattern of stagnation is still predictable if count the failed ones. Because when products fail to receive initial approval, firms are likely to face greater difficulty in deepening or diversifying the product portfolios.^30,31

The transition structure further suggests that most enterprises have failed to or decided not to expand their product portfolio, as 193 enterprises stay in State 1. Even for those 76 enterprises in State 2 who deepened their product portfolio in an identical application area, further expansion to multiple application areas was less frequent.76 enterprises (26%) remained in State 2 with repeated submissions in the same application area, while only 5 of the 81 enterprises entering State 2 (6%) transitioned to State 3 (Figure 2). Enterprises often have economic and strategic incentives to expand their product portfolios, as portfolio expansion may enable them to exploit economies of scope, leverage underutilized resources, respond to heterogeneous market demand, increase market share, and support firm growth. ³² It’s same in healthcare. Euh and Lee (2021) found that pharmaceutical companies diversifying into medical devices achieved higher long-term growth potential (meta-technology ratio: 94.54%) than non-diversified firms (80.39%), despite lower current operational efficiency. ³³ The innovation trajectory found contradicts the conventional perception in product portfolio building. In the context of SaMD, the stagnating pattern reflects a voluntary or forced interruption in the innovation trajectory, preventing most enterprises from reaching the stage at which the benefits of product portfolio expansion can materialize.

Heterogeneity in innovation trajectory

Heterogeneity in innovation trajectories revealed distinct situation faced by enterprises. In the Cox PH regression, large enterprises demonstrated significantly higher numbers of FDA approvals and broader application area coverage than their smaller counterparts. This suggests that larger enterprises have a greater capacity to navigate complex regulatory processes and engage in innovation diversification, capabilities which are essential for success in digital health ecosystems characterized by high institutional and technological complexity. ³⁴ This observation is consistent with previous research highlighting the importance of organizational resources and dynamic capabilities in facilitating sustained digital innovation.^5,35–37 Industry classification did not exhibit statistically significant effects for most industry types; however, software enterprises showed a significantly higher expansion hazard (HR=2.15, $p$ =0.04), though this effect was absent in the focused-repeat pathway and driven by diversification. These results imply that legacy sectoral expertise does not strongly influence innovation outcomes in the SaMD field. This finding aligns with existing ecosystem literature that underscores the diminishing significance of traditional industry boundaries in digital innovation.^38–41 In the healthcare, innovation heavily relies on data integration, platform interoperability, and adaptive regulatory navigation, capabilities which are usually associated with enterprise scale. As digital innovation results from the recombination of diversified digital resources,^4,42 the innovation trajectory of SaMD in macro is shaped by enterprise-specific configurations rather than industry dynamics.

Conversely, in specialized innovation contexts, where enterprises pursued a narrow yet deep trajectory, observable structural characteristics offer less explanatory power. When we examined State 2 enterprises have multiple FDA-approved SaMD within a single application area, the effects of both the size effect was attenuated. Only micro enterprises retained statistical significance (HR=0.21, $p$ <0.005), while medium and small enterprises showed no significant difference from large enterprises. Software showed significance only in the overall model (HR=2.15, $p$ =0.04), not in the cause-specific focused-repeat model, indicating that the effect is driven by diversification rather than depth. Enterprise size primarily determines whether an enterprise can make a second submission, rather than deepening or diversifying.

The finding aligns with recent arguments that in digital innovation ecosystems, enterprises may follow unique learning and engagement paths that reflect their relational positioning, experiential learning, or regulatory co-creation practices.^31,43 Therefore, while structural attributes like size can explain macro-level patterns of innovation breadth and intensity, a detailed understanding of innovation trajectory requires examining micro factors like managerial cognition, regulatory engagement routines, and interorganizational networks. The finding also suggests that small and micro enterprises do not necessarily face systematic disadvantages compared to large incumbents in SaMD. In emerging technological fields, firm-size differences shape innovation trajectories. Large incumbents often capture early growth and commercialization advantages when the new field requires scale, R&D resources, complementary assets, data, regulatory capability, or ecosystem access.^44–46 However, this advantage is not deterministic. SMEs and startups can pioneer radical or disruptive product innovation, especially when the technological discontinuity is competence-destroying or architectural.^47–49 SME’s catch-up and scaling depend on building technological capabilities and accessing complementary assets through alliances, data, regulatory expertise, and commercialization partners.^45,50,51 In the context of SaMD, the precise nature of these compensatory mechanisms remains unclear in the present study due to limitations of data.

Barrier to innovation of SaMD

The current SaMD landscape exhibits pronounced concentration in specific medical specialties, notably medical imaging and radiology, where structured data availability facilitates innovation.^25,52,53 This concentration is not coincidental but stems from the inherent dependencies on specialized resources for digital health innovation. SaMD functionality relies on high-quality data and domain-specific knowledge and clinical validation opportunities, creating a structural innovation constraint where resource accessibility becomes the primary determinant of innovation distribution across medical specialties.^52,54–56 The multi-state analysis demonstrated this constraint empirically: 55% (11 of 20) of enterprises reaching advanced innovation states (States 3 and 4) were large enterprises, despite large enterprises accounting for only 21% (n=60) of the total sample. This overrepresentation indicates that advancing beyond initial innovation stages requires the established resource infrastructure that large enterprises possess, while potentially transformative applications in areas like digital therapeutics and remote monitoring remained significantly underrepresented despite their potential clinical impact.^57,58 These findings revealed a barrier that shapes digital innovation trajectories in regulated healthcare markets.

Our empirical findings offer a fresh perspective on understanding the innovation of SaMD in the healthcare sector. The results uncover a distinctly stagnating pattern marked by significant stagnation in SaMD. 93.1% (n=269) of the SaMD developer enterprises remain at initial innovation stages (States 1, State 2), showing limited advancement beyond their first or second regulatory approval. While new entrants don’t necessarily need to diversify at the initial stage,^40,59,60 prolonged single-product stagnation contradicts patterns of successful digital healthcare technology ventures.^40,50,61 This concern intensifies in comparison to enterprises in State 2, which expand portfolios even in identical application areas. In general, single-product portfolios expose firms to market concentration risk, technological obsolescence, and limited growth potential. ⁶² In rapidly evolving markets, portfolio breadth reduces the hazard of exit from the market. ⁶³ The observed pattern in SaMD indicates that either strategic constraints or fundamental scaling challenges exist.

We propose that the stagnation in SaMD arises from a structural innovation barrier that is uniquely present in regulated market contexts, creating significant obstacles to sustained innovation trajectories (Figure 7). While our analysis captures enterprise-level outcomes, the underlying mechanisms extend beyond firm boundaries. Recent evidence from the EU context confirms that regulatory complexity, clinical data access, and partnership networks constitute ecosystem-level barriers that disproportionately constrain smaller firms SaMD startups face particular obstacles conducting clinical investigations due to financial costs and limited access to existing clinical evidence, while navigating regulatory requirements demands specialized expertise that smaller organizations often lack. ⁵⁰ The size-dependent expansion patterns observed in our data (micro HR=0.19 vs. large) are consistent with these ecosystem-level constraints. Contrary to the continuous progression and portfolio expansion expected by traditional innovation models, enterprises fail to sustain momentum beyond initial innovation in SaMD. This innovation barrier, akin to the ‘valley of death’, becomes evident as enterprises strive to advance beyond initial innovations, necessitating increasingly specialized resources that are disproportionately challenging to access within regulatory constraints. In the SaMD context, Real-World Evidence (RWE), the clinical evidence derived from the actual usage of medical products, is critical for product innovation and following FDA approval application in US. A similar regulatory burden is posed on EU enterprises as well. Clinical evaluation and postmarket surveillance are key requirements of the Regulation (EU) 2017/745 also known as the MDR (Medical Device Regulation) in EU. ⁵⁰ Accessing such evidence requires high-barrier assets, including factors like direct clinical access, specialized medical datasets, and domain-specific knowledge, alongside the significant funding required for clinical trials and regulatory navigation of new applications. ⁵⁰ The presence of barriers hinders small or micro-enterprises from achieving sustainable innovation gains, while only large enterprises can continue product development. This phenomenon explains why 93% (n=269) of enterprises fail to diversify beyond their initial application area despite market opportunities. Healthcare regulations also impose unique requirements for clinical validation, compliance documentation, and ongoing surveillance, significantly increasing resource intensity for each subsequent innovation.^58,64

Figure 7.

Schematic drawing of the development stages and innovation barriers of SaMD.

In digital health, new entrants, particularly startups that cannot access external specialized resources, produce more disruptive innovation.^25,65 Conclusively, this firm-level findings carry implications for the innovation systems perspective. The strong size-dependence of portfolio expansion, combined with the limited and pathway-specific nature of industry effects, suggests that the relevant constraints operate not at the sectoral level but at the interface between firms and the regulatory and clinical infrastructure they depend on. We propose that the sustainable innovation of SaMD critically depends on the development of an innovation system where enterprises can orchestrate resources across boundaries and necessitate alliance capabilities beyond traditional R&D expertise.^66,67 Such a system manifests as collaborative development platforms led by public hospitals or health systems, which serve as “resource hubs” providing innovators with access to RWE. This orchestration would facilitate the exchange of critical assets, including the large-scale datasets required for product iteration and direct clinical trial accessibility, thereby lowering the structural barriers for enterprises to maintain sustained innovation trajectories. An emerging example is Juntendo University’s GAUDI (Global Alliance Under the Dynamic Innovation) program in Japan, which leverages a clinical infrastructure spanning 6 affiliated hospitals and over 3 million annual outpatient visits to provide SaMD developers with access to clinical trial support, real-world evidence, and domain expertise. By functioning as a resource hub bridging clinical institutions and technology innovators, programs like GAUDI illustrate how academic health systems can lower structural barriers for enterprises that lack independent access to clinical validation resources. Future research incorporating qualitative case studies of such collaborative ecosystems would further elucidate the mechanisms through which they facilitate sustained innovation.

Limitations and future research

This study has several limitations that warrant consideration. First, the dataset included only U.S. FDA-approved SaMD products and their applicant enterprises covering the period January 2012 to February 2022, from which we were unable to distinguish enterprises by like investment capacity, strategic focus. Also, all U.S. FDA-approved SaMD products have been successful in product innovation. Our dataset does not include the failed innovation attempts. Moreover, regulatory frameworks differ across jurisdictions. The European Medical Device Regulation requires conformity assessment through notified bodies, while Japan’s PMDA follows a distinct pre-market approval pathway. Whether the stagnation and size-dependent patterns observed here generalize beyond the US FDA context remains an open question for future comparative research. Second, identifying SaMD products within the FDA database presents methodological challenges due to the absence of an official filtering mechanism. While our keyword-based extraction approach yielded a dataset between January 2012 to February 2022, alternative filtering methodologies might reveal additional patterns or insights. Third, we proposed several potential factors while not specifying the factors produced the innovation barrier. However, this study did not conduct a causal assessment to verify and evaluate which factor influences the result of digital innovation due to the limitations of the dataset. Future research should examine the causal relationships between potential determinants and firm performance in SaMD innovation. Case studies or interviews with enterprise leaders could empirically validate the links between these factors and the proposed barriers. Further investigation of technological advances, like AI and IoT integration, would also deepen understanding of how innovation trajectories evolve. Also, studies on how regulatory frameworks in other jurisdictions like EU, Japan impact innovation trajectories provide comparative insights.

Conclusions

This study advances the understanding of how enterprises navigate product development of SaMD through the multi-state modelling of 289 SaMD enterprises. The analysis reveals distinctly stagnating innovation trajectories characterized by heterogeneous diversification patterns across industry classifications. Enterprises demonstrate a significantly lower propensity for product portfolio diversification than traditionally theorized, with 93% (n=269) remaining in single-application states throughout the observation period. This diverges from conventional assumptions about the diffusion of innovation. When stratified by enterprise characteristics, large enterprises continuously pursued product development and diversification of product portfolio, while small and micro enterprises faced substantial progression barriers in deepening or diversifying after their initial innovation. The differential in progression rates between enterprise types revealed a fundamental resource mismatch where start-up firms, mostly leading disruptive innovation, face the greatest resource constraints. To resolve this issue, an innovation system that facilitates resource orchestration under highly regulated conditions is essential.

Supplemental material

Supplemental material - Innovation trajectory of software as a medical device: Evidence from the US FDA-approved products

Supplemental material for Innovation trajectory of software as a medical device: Evidence from the US FDA-approved products by Jiakan Yu, Jiajie, Zhang, Shuto Miyashita, and Sengoku Shintaro in DIGITAL HEALTH.