Trends of Gaps Between Health-Adjusted Life Expectancy and Life Expectancy at the Regional Level in Korea Using a Group-Based Multi-Trajectory Modeling Approach (2008-2019).

BACKGROUND: Health-adjusted life expectancy (HALE) is an indicator of the average lifespan in good health. Through this study, we aimed to identify regional disparities in the gap between HALE and life expectancy, considering the trends that have changed over time in Korea. METHODS: We employed a group-based multi-trajectory modeling approach to capture trends in the gap between HALE and life expectancy at the regional level from 2008 to 2019. HALE was calculated using incidence-based "years lived with disability." This methodology was also employed in the Korean National Burden of Disease Study. RESULTS: Based on five different information criteria, the most fitted number of trajectory groups was seven, with at least 11 regions in each group. Among the seven groups, one had an exceptionally large gap between HALE and life expectancy compared to that of the others. This group was assigned to 17 regions, of which six were metropolitan cities. CONCLUSION: Based on the results of this study, we identified regions in which health levels have deteriorated over time, particularly within specific areas of metropolitan cities. These findings can be used to design comprehensive policy interventions for community health promotion and urban regeneration projects in the future.

Establishment of a Dataset for the Traditional Korean Medicine Examination in Healthy Adults.

We established a protocol for the traditional Korean medicine examination (KME) and methodically gathered data following this protocol. Potential indicators for KME were extracted through a literature review; the first KME protocol was developed based on three rounds of expert opinions. The first KME protocol's feasibility was confirmed, and data were collected over four years from traditional Korean medicine (KM) hospitals, focusing on healthy adults, using the final KME protocol. A literature review identified 175 potential core indicators, condensed into 73 indicators after three rounds of expert consultation. The first KME protocol, which was categorized under questionnaires and medical examinations, was developed after the third round of expert opinions. A pilot study using the first KME protocol was conducted to ensure its validity, leading to modifications resulting in the development of the final KME protocol. Over four years, data were collected from six KM hospitals, focusing on healthy adults; we obtained a dataset comprising 11,036 healthy adults. This is the first protocol incorporating core indicators of KME in a quantitative form and systematically collecting data. Our protocol holds potential merit in evaluating predisposition to diseases or predicting diseases.

Effectiveness of calcipotriol/betamethasone dipropionate aerosol foam in patients with small versus large plaque psoriasis in routine practice in South Korea.

Small plaque psoriasis is the typical form of chronic plaque psoriasis affecting adults in South Korea. The effectiveness of calcipotriol/betamethasone dipropionate (Cal/BD) aerosol foam for large and small psoriasis plaques has not previously been examined. We performed a post hoc analysis of a recent, 4-week observational study of Cal/BD aerosol foam use in routine clinical practice in South Korea. Investigator Global Assessment response ([IGA] 0/1 at week 4), Patient Global Assessment response ([PaGA] 0/1 at week 4), change in Psoriasis Area and Severity Index (PASI), changes in psoriasis symptom scores, change in the Dermatology Life Quality Index (DLQI), and the proportion of patients achieving DLQI ≤5 were analyzed for patients with small (≤5 cm; n = 131) or large (>5 cm; n = 35) baseline plaque size. IGA response rates were similar for patients with small and large plaques (59.5% and 51.4% respectively). Similarly, there was no significant difference between the small and large groups in mean change in PASI (-2.20 vs -3.34), the proportions of patients with DLQI ≤5 (62.3% vs 54.3%) or PaGA 0/1 (29.2% vs 40.0%). Mean improvements in DLQI (-4.04 vs -6.20) and in psoriasis symptoms including itching (-1.50 vs -2.83), sleep loss (-0.67 vs -1.89), dryness (-1.57 vs -2.97), scaling (-1.21 vs -3.57), and redness (-1.17 vs -3.11) were greater in patients with large plaques than those with small plaques. Itching and DLQI differences were not statistically significant after adjustment for baseline characteristics. Stratification by body surface area affected eliminated statistically significant differences between the groups for most outcomes. In conclusion, this analysis suggests that Cal/BD aerosol foam is an effective, well-accepted treatment for adult patients with the small plaques typical of chronic plaque psoriasis in South Korea, as well as for those with large plaques.

Vaccination and the risk of post-acute sequelae after COVID-19 in the Omicron-predominant period.

OBJECTIVES: To assess the association of primary and third doses of vaccination with the risk of post-acute sequelae of COVID-19. METHODS: This retrospective cohort study utilized a combined database of nationwide health care claims data, COVID-19 patient registry, and vaccination records from South Korea. Individuals diagnosed with COVID-19 in the Omicron variant-dominant period of January-March 2022 were tracked for 30-120 days post-infection. The exposure of interest was the receipt of primary and third doses of the SARS-CoV-2 vaccine. The occurrence of 26 specific conditions in eight domains was compared using Cox regression with inverse probability of treatment weighting. RESULTS: This study included 394 773 unvaccinated individuals and 7 604 081 individuals receiving ≥2 doses of vaccine. Compared with unvaccinated individuals, vaccination with at least two doses was associated with a reduced risk (adjusted hazard ratio; 95% CI) of several conditions, including ischaemic heart disease (0.73; 0.57-0.94), heart failure (0.55; 0.48-0.63), cardiac dysrhythmia (0.72; 0.61-0.85), cardiac arrest (0.41; 0.33-0.51), pulmonary embolism (0.66; 0.52-0.84), venous thromboembolism (0.54; 0.44-0.66), acute renal failure (0.56; 0.46-0.67), new dialysis (0.45; 0.34-0.59), chronic obstructive pulmonary disease (0.74; 0.65-0.84), acute pancreatitis (0.64; 0.51-0.80), and diabetes (0.82; 0.71-0.95). The risks of heart failure, cardiac dysrhythmias, cardiac arrest, pulmonary embolism, and new dialysis were lower in those who were vaccinated with three doses compared with those who were vaccinated with two doses. DISCUSSION: Vaccination was associated with a decreased risk of post-acute sequelae of COVID-19, suggesting its potential role in mitigating the indirect impacts of COVID-19.

Rethinking skip connections in Spiking Neural Networks with Time-To-First-Spike coding.

Time-To-First-Spike (TTFS) coding in Spiking Neural Networks (SNNs) offers significant advantages in terms of energy efficiency, closely mimicking the behavior of biological neurons. In this work, we delve into the role of skip connections, a widely used concept in Artificial Neural Networks (ANNs), within the domain of SNNs with TTFS coding. Our focus is on two distinct types of skip connection architectures: (1) addition-based skip connections, and (2) concatenation-based skip connections. We find that addition-based skip connections introduce an additional delay in terms of spike timing. On the other hand, concatenation-based skip connections circumvent this delay but produce time gaps between after-convolution and skip connection paths, thereby restricting the effective mixing of information from these two paths. To mitigate these issues, we propose a novel approach involving a learnable delay for skip connections in the concatenation-based skip connection architecture. This approach successfully bridges the time gap between the convolutional and skip branches, facilitating improved information mixing. We conduct experiments on public datasets including MNIST and Fashion-MNIST, illustrating the advantage of the skip connection in TTFS coding architectures. Additionally, we demonstrate the applicability of TTFS coding on beyond image recognition tasks and extend it to scientific machine-learning tasks, broadening the potential uses of SNNs.

Measuring the Burden of Disease in Korea Using Disability-Adjusted Life Years (2008-2020).

BACKGROUND: The measurement of health levels and monitoring of characteristics and trends among populations and subgroups are essential for informing evidence-based policy decisions. This study aimed to examine the burden of disease in Korea for both the total population and subgroups in 2020, as well as analyze changes in disease burden from 2008 to 2020. METHODS: We employed the methodology developed in the Korean National Burden of Disease and Injuries Study to calculate disability-adjusted life years (DALYs) by sex, causes, region, and income level from 2008 to 2020. DALYs were derived by combining years of life lost and years lived with disability. RESULTS: In 2020, the burden of disease for the Korean population was estimated to be 25,439 DALYs per 100,000 population, reflecting a 13.8% increase since 2008. The leading causes of DALYs were diabetes mellitus, followed by low back pain and ischemic stroke. A sex-specific gap reversal was observed, with the disease burden for men surpassing that of women starting in 2017. Furthermore, variations in disease burden were identified across 250 regions and income quintiles. CONCLUSION: It is imperative to establish appropriate health policies that prioritize the diseases with significantly increasing burdens and subgroups experiencing high disease burdens. The findings of this study are expected to serve as a foundation for developing healthcare policies aimed at improving the health levels of Koreans and achieving health equity.

Progranulin haploinsufficiency mediates cytoplasmic TDP-43 aggregation with lysosomal abnormalities in human microglia.

BACKGROUND: Progranulin (PGRN) haploinsufficiency due to progranulin gene (GRN) variants can cause frontotemporal dementia (FTD) with aberrant TAR DNA-binding protein 43 (TDP-43) accumulation. Despite microglial burden with TDP-43-related pathophysiology, direct microglial TDP-43 pathology has not been clarified yet, only emphasized in neuronal pathology. Thus, the objective of this study was to investigate TDP-43 pathology in microglia of patients with PGRN haploinsufficiency. METHODS: To design a human microglial cell model with PGRN haploinsufficiency, monocyte-derived microglia (iMGs) were generated from FTD-GRN patients carrying pathogenic or likely pathogenic variants (p.M1? and p.W147*) and three healthy controls. RESULTS: iMGs from FTD-GRN patients with PGRN deficiency exhibited severe neuroinflammation phenotype and failure to maintain their homeostatic molecular signatures, along with impaired phagocytosis. In FTD-GRN patients-derived iMGs, significant cytoplasmic TDP-43 aggregation and accumulation of lipid droplets with profound lysosomal abnormalities were observed. These pathomechanisms were mediated by complement C1q activation and upregulation of pro-inflammatory cytokines. CONCLUSIONS: Our study provides considerable cellular and molecular evidence that loss-of-function variants of GRN in human microglia can cause microglial dysfunction with abnormal TDP-43 aggregation induced by inflammatory milieu as well as the impaired lysosome. Elucidating the role of microglial TDP-43 pathology in intensifying neuroinflammation in individuals with FTD due to PGRN deficiency and examining consequential effects on microglial dysfunction might yield novel insights into the mechanisms underlying FTD and neurodegenerative disorders.

Trends in Healthy Life Expectancy (HALE) and Disparities by Income and Region in Korea (2008-2020): Analysis of a Nationwide Claims Database.

BACKGROUND: Healthy life expectancy is a well-recognized indicator for establishing health policy goals used in Korea's Health Plan. This study aimed to explore Koreans' healthy life expectancy and its gender, income, and regional disparities from 2008 to 2020. METHODS: This study was conducted on the entire population covered by health insurance and medical aid program in Korea. The incidence-based "years lived with disability" for 260 disease groups by gender, income level, and region was calculated employing the methodology developed in the Korean National Burden of Disease Study, and it was used as the number of healthy years lost to calculate health-adjusted life expectancy (HALE). RESULTS: Koreans' HALE increased from 68.89 years in 2008 to 71.82 years in 2020. Although the gender disparity in HALE had been decreasing, it increased to 4.55 years in 2020. As of 2020, 5.90 years out of 8.67 years of the income disparity (Q5-Q1) in HALE were due to the disparity between Q1 and Q2, the low-income groups. Income and regional disparities in HALE exhibited an increasing trend, and these disparities were higher in men than in women. CONCLUSION: A subgroup with a low health level was identified through the HALE results, and it was confirmed that improving the health level of this population can reduce health inequalities and improve health at the national level. Further exploration of the HALE calculation methodology may help in the development of effective policies such as prioritizing interventions for health risk factors.

Microneedle patch-based enzyme-linked immunosorbent assay to quantify protein biomarkers of tuberculosis.

There is a clinical need for differential diagnosis of the latent versus active stages of tuberculosis (TB) disease by a simple-to-administer test. Alpha-crystallin (Acr) and early secretory antigenic target-6 (ESAT-6) are protein biomarkers associated with the latent and active stages of TB, respectively, and could be used for differential diagnosis. We therefore developed a microneedle patch (MNP) designed for application to the skin to quantify Acr and ESAT-6 in dermal interstitial fluid by enzyme-linked immunosorbent assay (ELISA). We fabricated mechanically strong microneedles made of polystyrene and coated them with capture antibodies against Acr and ESAT-6. We then optimized assay sensitivity to achieve a limit of detection of 750 pg/ml and 3,020 pg/ml for Acr and ESAT-6, respectively. This study demonstrates the feasibility of an MNP-based ELISA for differential diagnosis of latent TB disease.

Rapid assays of SARS-CoV-2 virus and noble biosensors by nanomaterials.

The COVID-19 outbreak caused by SARS-CoV-2 in late 2019 has spread rapidly across the world to form a global epidemic of respiratory infectious diseases. Increased investigations on diagnostic tools are currently implemented to assist rapid identification of the virus because mass and rapid diagnosis might be the best way to prevent the outbreak of the virus. This critical review discusses the detection principles, fabrication techniques, and applications on the rapid detection of SARS-CoV-2 with three categories: rapid nuclear acid augmentation test, rapid immunoassay test and biosensors. Special efforts were put on enhancement of nanomaterials on biosensors for rapid, sensitive, and low-cost diagnostics of SARS-CoV-2 virus. Future developments are suggested regarding potential candidates in hospitals, clinics and laboratories for control and prevention of large-scale epidemic.

Continued JAK inhibitor treatment on the risk of recurrent herpes zoster reactivation in patients with immune-mediated inflammatory diseases: A nationwide population-based study in South Korea.

BACKGROUND: To investigate the risk of recurrent herpes zoster (HZ) reactivation under continued Janus kinase inhibitor (JAKi) therapy in patients with immune-mediated inflammatory diseases (IMID) who developed HZ reactivation. METHODS: Data from the Korean Health Insurance Review and Assessment Service (HIRA) of patients with rheumatoid arthritis (RA) or ulcerative colitis (UC) gathered from 2007 to 2021 were analyzed. RESULTS: A total of 3947 (RA 3540, UC 407) receiving JAKi were included. After median 0.95 years (IQR, 0.93-2.58) of therapy, 611 (15.5%) patients developed HZ reactivation (incidence rate: 8.38/100 person-years [PY]). After excluding 151 patients with lack of data after HZ reactivation, 460 patients (JAKi continuation group, n = 386 [83.9%]; JAKi discontinuation group, n = 74 [16.1%]) were analyzed for the risk of subsequent recurrent HZ reactivation. During further follow-up of median 1.11 years (IQR, 0.53-1.91), 36 (9.3%) and 6 (8.1%) patients in the JAKi continuation group and JAKi discontinuation group experienced a recurrence of HZ, respectively. The incidence rate of subsequent recurrent HZ reactivation was not significantly different between the two groups (5.3/100 vs. 5.9/100 PY; P = 0.52). After adjusting for age, sex, usage of corticosteroids, and antiviral agents, continued use of JAKi was not a significant risk factor for subsequent HZ reactivation (adjusted hazard ratio, 0.71 [CI, 0.29-1.72], P = 0.45). CONCLUSION: In this nationwide population-based study on patients with RA or UC, continued use of JAKi was not associated with a significant risk of subsequent recurrent HZ reactivation. JAKi therapy may be maintained in patients with IMID even after HZ reactivation.

Association of Muscle Mass Loss with Diabetes Development in Liver Transplantation Recipients.

BACKGRUOUND: Post-transplant diabetes mellitus (PTDM) is one of the most significant complications after transplantation. Patients with end-stage liver diseases requiring transplantation are prone to sarcopenia, but the association between sarcopenia and PTDM remains to be elucidated. We aimed to investigate the effect of postoperative muscle mass loss on PTDM development. METHODS: A total of 500 patients who underwent liver transplantation at a tertiary care hospital between 2005 and 2020 were included. Skeletal muscle area at the level of the L3-L5 vertebrae was measured using computed tomography scans performed before and 1 year after the transplantation. The associations between the change in the muscle area after the transplantation and the incidence of PTDM was investigated using a Cox proportional hazard model. RESULTS: During the follow-up period (median, 4.9 years), PTDM occurred in 165 patients (33%). The muscle mass loss was greater in patients who developed PTDM than in those without PTDM. Muscle depletion significantly increased risk of developing PTDM after adjustment for other confounding factors (hazard ratio, 1.50; 95% confidence interval, 1.23 to 1.84; P=0.001). Of the 357 subjects who had muscle mass loss, 124 (34.7%) developed PTDM, whereas of the 143 patients in the muscle mass maintenance group, 41 (28.7%) developed PTDM. The cumulative incidence of PTDM was significantly higher in patients with muscle loss than in patients without muscle loss (P=0.034). CONCLUSION: Muscle depletion after liver transplantation is associated with increased risk of PTDM development.

Febuxostat dose requirement according to renal function in patients who achieve target serum urate levels: A retrospective cohort study.

OBJECTIVES: To determine the febuxostat dose requirement according to renal function in patients who achieve target serum urate (SU) levels. METHODS: Of 3153 gout patients who underwent febuxostat treatment, 873 patients with an initial SU level>6mg/dL were included and categorized by the estimated glomerular filtration rate: normal, chronic kidney disease (CKD) stage 3, and stages 4-5. Ninety-five patients with insufficient follow-up were further excluded. The dose of febuxostat in patients who achieved the SU target (< 6mg/dL) was defined as the average daily dosage at the time of SU target achievement. RESULTS: The cohort of 778 gout patients had a median age of 52.0 years (IQR, 41.0-63.0) and comprised 711 (91.4%) men. The mean SU at febuxostat initiation was higher in the CKD 4-5 (9.6 [± 3.1] mg/dL) than in the other groups (CKD 3, 8.7 [± 1.7]; normal, 8.4 [± 1.7]; P<0.001). Patients achieved target SU at a median of 4.0 (1.9-9.6) months and in those who achieved target SU, the dose of febuxostat at the time of SU target achievement was significantly lower in the CKD 4-5 group (50.0 [± 16.5] mg) than in the other groups (vs. CKD stage 3, 60.0 [± 19.5] mg; P<0.01, vs. normal, 60.0 [± 19.8] mg; P<0.01). Furthermore, CKD stage 4-5 had a negative correlation with the febuxostat dose requirement (Beta: -2.334, P<0.05). CONCLUSION: Among patients who achieved SU target, those with severely decreased renal function (CKD 4-5) required a lower febuxostat dose to achieve the target SU level compared to patients with normal or mild renal impairment.

Risk of Severe COVID-19 and Protective Effectiveness of Vaccination Among Solid Organ Transplant Recipients.

BACKGROUND: Solid organ transplant recipients (SOTRs) are at higher risk for severe infection. However, the risk for severe COVID-19 and vaccine effectiveness among SOTRs remain unclear. METHODS: This retrospective study used a nationwide health care claims database and COVID-19 registry from the Republic of Korea (2020 to 2022). Adult SOTRs diagnosed with COVID-19 were matched with up to 4 non-SOTR COVID-19 patients by propensity score. Severe COVID-19 was defined as treatment with high-flow nasal cannulae, mechanical ventilation, or extracorporeal membrane oxygenation. RESULTS: Among 6783 SOTRs with COVID-19, severe COVID-19 was reported with the highest rate in lung transplant recipients (13.16%), followed by the heart (6.30%), kidney (3.90%), and liver (2.40%). SOTRs had a higher risk of severe COVID-19 compared to non-SOTRs, and lung transplant recipients showed the highest risk (adjusted odds ratio, 18.14; 95% confidence interval [CI], 8.53-38.58). Vaccine effectiveness against severe disease among SOTRs was 47% (95% CI, 18%-65%), 64% (95% CI, 49%-75%), and 64% (95% CI, 29%-81%) for 2, 3, and 4 doses, respectively. CONCLUSIONS: SOTRs are at significantly higher risk for severe COVID-19 compared to non-SOTRs. Vaccination is effective in preventing the progression to severe COVID-19. Efforts should be made to improve vaccine uptake among SOTRs, while additional protective measures should be developed.

Comparative safety analysis of mRNA and adenoviral vector COVID-19 vaccines: a nationwide cohort study using an emulated target trial approach.

OBJECTIVE: This nationwide cohort study compared the incidence of adverse events of special interest (AESIs) between adenoviral vector-based (ChAdOx1) and mRNA-based (BNT162b2 or mRNA-1273) coronavirus disease 2019 (COVID-19) vaccines. METHODS: A targeted trial emulation study was conducted using data from the National Health Insurance Service database. Vaccinees aged 18-85 years who had received at least one dose of ChAdOx1 or an mRNA-based vaccine were identified. The 42-day risks of AESIs were calculated. RESULTS: A total of 1 767 539 ChAdOx1 vaccinees were matched exactly with mRNA vaccinees according to their risk factors. The 42-day risks of adverse events were low (∼0 to 176 events per 100 000 persons in both vaccine groups), and the incidence rates of AESIs were comparable between the two platforms, except for a higher occurrence of acute cardiac injury (incidence rate ratio [IRR], 1.22; 95% CI, 1.10-1.35), myocarditis or pericarditis (IRR, 2.14; 95% CI, 1.14-4.04), and arrhythmia (IRR, 1.46; 95% CI, 1.24-1.71) in mRNA vaccinees. The incidence of Guillain-Barré syndrome (IRR, 0.20; 95% CI, 0.06-0.69), vasovagal syncope (IRR, 0.77; 95% CI, 0.62-0.97), radiculopathy (IRR = 0.59, 95% CI, 0.41-0.84), and aseptic arthritis (IRR, 0.81; 95% CI, 0.70-0.93) was significantly lower in mRNA-based vaccinees compared with ChAdOx1 vaccinees. DISCUSSION: A remarkable platform-dependent difference was observed in the safety profiles of COVID-19 vaccines, particularly for myocarditis or pericarditis and Guillain-Barré syndrome. However, the overall risk of AESIs was low for both vaccine platforms.

Nationwide Trends in Non-COVID-19 Infectious Disease Laboratory Tests in the Era of the COVID-19 Pandemic in Korea.

The coronavirus disease 2019 pandemic has brought significant changes to infectious disease management globally. This study explored changes in clinical microbiological testing trends and their implications for infectious disease incidence and medical utilization during the pandemic. We collected nationwide claims for monthly clinical microbiology tests from January 2018 to March 2022 using the National Health Insurance Service database. Seasonal autoregressive integrated moving average models were employed to make predictions for each disease based on the baseline period (January 2018 to January 2020). The results showed a significant decrease in general bacterial and fungal cultures, respiratory infectious disease-related, and inflammatory markers, while the representatives of tests for vector-borne diseases, healthcare-associated infections, and chronic viral infections remained stable. The study highlights the potential of clinical microbiological testing trends as an additional surveillance tool and offers implications for future infectious disease management and surveillance strategies in pandemic settings.

An analysis of variants in TARDBP in the Korean population with amyotrophic lateral sclerosis: comparison with previous data.

The TARDBP gene variant is a known major cause of amyotrophic lateral sclerosis (ALS), with limited reports of Korean patients with ALS harboring the variants in TARDBP. This large cohort study introduces four ALS patients who share the p.M337V variant of the TARDBP, allowing for an investigation of clinical characteristics and prognosis by analyzing previously reported cases with the same variant. From November 2014 to August 2022, participants were recruited from two tertiary hospitals in Seoul, Korea. Clinical characteristics of patients diagnosed with ALS carrying the variant in TARDBP were evaluated. Previous articles demonstrating subjects' characteristics were reviewed. Four patients were identified with the pathogenic missense variant (c.1009A>G; p.M337V) in the TARDBP. The mean age of onset was 55 years old, and none of the patients showed severe cognitive impairment. Sixty-three patients carrying the p.M337V variant in TARDBP from this study and previous reports delineated young age of onset (51.6 years), high frequency of bulbar onset patients (61.9%), and low comorbidity of frontotemporal dementia. This study reveals the presence of pathogenic variant of TARDBP in Korea and emphasizes the importance of genetic screening of the TARDBP gene, in diagnosing ALS and evaluating prognosis among familial and simplex ALS patients in Korea.

Efficient human activity recognition with spatio-temporal spiking neural networks.

In this study, we explore Human Activity Recognition (HAR), a task that aims to predict individuals' daily activities utilizing time series data obtained from wearable sensors for health-related applications. Although recent research has predominantly employed end-to-end Artificial Neural Networks (ANNs) for feature extraction and classification in HAR, these approaches impose a substantial computational load on wearable devices and exhibit limitations in temporal feature extraction due to their activation functions. To address these challenges, we propose the application of Spiking Neural Networks (SNNs), an architecture inspired by the characteristics of biological neurons, to HAR tasks. SNNs accumulate input activation as presynaptic potential charges and generate a binary spike upon surpassing a predetermined threshold. This unique property facilitates spatio-temporal feature extraction and confers the advantage of low-power computation attributable to binary spikes. We conduct rigorous experiments on three distinct HAR datasets using SNNs, demonstrating that our approach attains competitive or superior performance relative to ANNs, while concurrently reducing energy consumption by up to 94%.

Divide-and-conquer the NAS puzzle in resource-constrained federated learning systems.

Federated Learning (FL) is a privacy-preserving distributed machine learning approach geared towards applications in edge devices. However, the problem of designing custom neural architectures in federated environments is not tackled from the perspective of overall system efficiency. In this paper, we propose DC-NAS-a divide-and-conquer approach that performs supernet-based Neural Architecture Search (NAS) in a federated system by systematically sampling the search space. We propose a novel diversified sampling strategy that balances exploration and exploitation of the search space by initially maximizing the distance between the samples and progressively shrinking this distance as the training progresses. We then perform channel pruning to reduce the training complexity at the devices further. We show that our approach outperforms several sampling strategies including Hadamard sampling, where the samples are maximally separated. We evaluate our method on the CIFAR10, CIFAR100, EMNIST, and TinyImagenet benchmarks and show a comprehensive analysis of different aspects of federated learning such as scalability, and non-IID data. DC-NAS achieves near iso-accuracy as compared to full-scale federated NAS with 50% fewer resources.