Artificial intelligence of arterial Doppler waveforms to predict major adverse outcomes among patients with diabetes mellitus.

OBJECTIVE: Patients with diabetes mellitus (DM) are at increased risk for peripheral artery disease (PAD) and its complications. Arterial calcification and non-compressibility may limit test interpretation in this population. Developing tools capable of identifying PAD and predicting major adverse cardiac event (MACE) and limb event (MALE) outcomes among patients with DM would be clinically useful. Deep neural network analysis of resting Doppler arterial waveforms was used to detect PAD among patients with DM and to identify those at greatest risk for major adverse outcome events. METHODS: Consecutive patients with DM undergoing lower limb arterial testing (April 1, 2015-December 30, 2020) were randomly allocated to training, validation, and testing subsets (60%, 20%, and 20%). Deep neural networks were trained on resting posterior tibial arterial Doppler waveforms to predict all-cause mortality, MACE, and MALE at 5 years using quartiles based on the distribution of the prediction score. RESULTS: Among 11,384 total patients, 4211 patients with DM met study criteria (mean age, 68.6 ± 11.9 years; 32.0% female). After allocating the training and validation subsets, the final test subset included 856 patients. During follow-up, there were 262 deaths, 319 MACE, and 99 MALE. Patients in the upper quartile of prediction based on deep neural network analysis of the posterior tibial artery waveform provided independent prediction of death (hazard ratio [HR], 3.58; 95% confidence interval [CI], 2.31-5.56), MACE (HR, 2.06; 95% CI, 1.49-2.91), and MALE (HR, 13.50; 95% CI, 5.83-31.27). CONCLUSIONS: An artificial intelligence enabled analysis of a resting Doppler arterial waveform permits identification of major adverse outcomes including all-cause mortality, MACE, and MALE among patients with DM.

Artificial intelligence for automatic detection of basal cell carcinoma from frozen tissue tangential biopsies.

Evaluation of basal cell carcinoma (BCC) involves tangential biopsies of a suspicious lesion that is sent for frozen sections and evaluated by a Mohs micrographic surgeon. Advances in artificial intelligence (AI) have made possible the development of sophisticated clinical decision support systems to provide real-time feedback to clinicians which could have a role in optimizing the diagnostic workup of BCC. There were 287 annotated whole-slide images of frozen sections from tangential biopsies, of which 121 contained BCC, that were used to train and test an AI pipeline to recognize BCC. Regions of interest were annotated by a senior dermatology resident, experienced dermatopathologist, and experienced Mohs surgeon, with concordance of annotations noted on final review. Final performance metrics included a sensitivity and specificity of 0.73 and 0.88, respectively. Our results on a relatively small dataset suggest the feasibility of developing an AI system to aid in the workup and management of BCC.

Artificial intelligence for the evaluation of peripheral artery disease using arterial Doppler waveforms to predict abnormal ankle-brachial index.

BACKGROUND: Patients with peripheral artery disease (PAD) are at increased risk for major adverse limb and cardiac events including mortality. Developing screening tools capable of accurate PAD identification is a necessary first step for strategies of adverse outcome prevention. This study aimed to determine whether machine analysis of a resting Doppler waveform using deep neural networks can accurately identify patients with PAD. METHODS: Consecutive patients (4/8/2015 - 12/31/2020) undergoing rest and postexercise ankle-brachial index (ABI) testing were included. Patients were randomly allocated to training, validation, and testing subsets (70%/15%/15%). Deep neural networks were trained on resting posterior tibial arterial Doppler waveforms to predict normal (> 0.9) or PAD (⩽ 0.9) using rest and postexercise ABI. A separate dataset of 151 patients who underwent testing during a period after the model had been created and validated (1/1/2021 - 3/31/2021) was used for secondary validation. Area under the receiver operating characteristic curves (AUC) were constructed to evaluate test performance. RESULTS: Among 11,748 total patients, 3432 patients met study criteria: 1941 with PAD (mean age 69 ± 12 years) and 1491 without PAD (64 ± 14 years). The predictive model with highest performance identified PAD with an AUC 0.94 (CI = 0.92-0.96), sensitivity 0.83, specificity 0.88, accuracy 0.85, and positive predictive value (PPV) 0.90. Results were similar for the validation dataset: AUC 0.94 (CI = 0.91-0.98), sensitivity 0.91, specificity 0.85, accuracy 0.89, and PPV 0.89 (postexercise ABI comparison). CONCLUSION: An artificial intelligence-enabled analysis of a resting Doppler arterial waveform permits identification of PAD at a clinically relevant performance level.

Deep learning for dermatologists: Part II. Current applications.

Because of a convergence of the availability of large data sets, graphics-specific computer hardware, and important theoretical advancements, artificial intelligence has recently contributed to dramatic progress in medicine. One type of artificial intelligence known as deep learning has been particularly impactful for medical image analysis. Deep learning applications have shown promising results in dermatology and other specialties, including radiology, cardiology, and ophthalmology. The modern clinician will benefit from an understanding of the basic features of deep learning to effectively use new applications and to better gauge their utility and limitations. In this second article of a 2-part series, we review the existing and emerging clinical applications of deep learning in dermatology and discuss future opportunities and limitations. Part 1 of this series offered an introduction to the basic concepts of deep learning to facilitate effective communication between clinicians and technical experts.

Deep learning for dermatologists: Part I. Fundamental concepts.

Artificial intelligence is generating substantial interest in the field of medicine. One form of artificial intelligence, deep learning, has led to rapid advances in automated image analysis. In 2017, an algorithm demonstrated the ability to diagnose certain skin cancers from clinical photographs with the accuracy of an expert dermatologist. Subsequently, deep learning has been applied to a range of dermatology applications. Although experts will never be replaced by artificial intelligence, it will certainly affect the specialty of dermatology. In this first article of a 2-part series, the basic concepts of deep learning will be reviewed with the goal of laying the groundwork for effective communication between clinicians and technical colleagues. In part 2 of the series, the clinical applications of deep learning in dermatology will be reviewed and limitations and opportunities will be considered.

Sequential organ failure assessment score improves survival prediction for left ventricular assist device recipients in intensive care.

BACKGROUND: Preoperative risk scores facilitate patient selection, but postoperative risk scores may offer valuable information for predicting outcomes. We hypothesized that the postoperative Sequential Organ Failure Assessment (SOFA) score would predict mortality after left ventricular assist device (LVAD) implantation. METHODS: We retrospectively reviewed data from 294 continuous-flow LVAD implantations performed at Mayo Clinic Rochester during 2007 to 2015. We calculated the EuroSCORE, HeartMate-II Risk Score, and RV Failure Risk Score from preoperative data and the APACHE III and Post Cardiac Surgery (POCAS) risk scores from postoperative data. Daily, maximum, and mean SOFA scores were calculated for the first 5 postoperative days. The area under receiver-operator characteristic curves (AUC) was calculated to compare the scoring systems' ability to predict 30-day, 90-day, and 1-year mortality. RESULTS: For the entire cohort, mortality was 5% at 30 days, 10% at 90 days, and 19% at 1 year. The Day 1 SOFA score had better discrimination for 30-day mortality (AUC 0.77) than the preoperative risk scores or the APACHE III and POCAS postoperative scores. The maximum SOFA score had the best discrimination for 30-day mortality (AUC 0.86), and the mean SOFA score had the best discrimination for 90-day mortality (AUC 0.82) and 1-year mortality (AUC 0.76). CONCLUSIONS: We observed that postoperative mean and maximum SOFA scores in LVAD recipients predict short-term and intermediate-term mortality better than preoperative risk scores do. However, because preoperative and postoperative risk scores each contribute unique information, they are best used in concert to predict outcomes after LVAD implantation.

Electrocardiogram screening for aortic valve stenosis using artificial intelligence.

AIMS: Early detection of aortic stenosis (AS) is becoming increasingly important with a better outcome after aortic valve replacement in asymptomatic severe AS patients and a poor outcome in moderate AS. We aimed to develop artificial intelligence-enabled electrocardiogram (AI-ECG) using a convolutional neural network to identify patients with moderate to severe AS. METHODS AND RESULTS: Between 1989 and 2019, 258 607 adults [mean age 63 ± 16.3 years; women 122 790 (48%)] with an echocardiography and an ECG performed within 180 days were identified from the Mayo Clinic database. Moderate to severe AS by echocardiography was present in 9723 (3.7%) patients. Artificial intelligence training was performed in 129 788 (50%), validation in 25 893 (10%), and testing in 102 926 (40%) randomly selected subjects. In the test group, the AI-ECG labelled 3833 (3.7%) patients as positive with the area under the curve (AUC) of 0.85. The sensitivity, specificity, and accuracy were 78%, 74%, and 74%, respectively. The sensitivity increased and the specificity decreased as age increased. Women had lower sensitivity but higher specificity compared with men at any age groups. The model performance increased when age and sex were added to the model (AUC 0.87), which further increased to 0.90 in patients without hypertension. Patients with false-positive AI-ECGs had twice the risk for developing moderate or severe AS in 15 years compared with true negative AI-ECGs (hazard ratio 2.18, 95% confidence interval 1.90-2.50). CONCLUSION: An AI-ECG can identify patients with moderate or severe AS and may serve as a powerful screening tool for AS in the community.

Red blood cell transfusion threshold and mortality in cardiac intensive care unit patients.

BACKGROUND: The benefit of red blood cell (RBC) transfusion in anemic critically-ill patients with cardiovascular disease is uncertain, as is the optimal threshold at which RBC transfusion should be considered. We sought to examine the association between RBC transfusion and mortality stratified by nadir Hgb level and admission diagnosis among cardiac intensive care unit (CICU) patients. METHODS: Retrospective single-center cohort of 11,754 CICU patients admitted between 2007 and 2018. The association between RBC transfusion and hospital mortality at each nadir Hgb (<8 g/dL, 8-9.9 g/dL, ≥10 g/dL) was assessed using multivariable logistic regression adjusted for the propensity to receive RBC transfusion. RESULTS: The study population had a mean age of 68±15 years, including 38% females; 1,134 (11.4%) received RBC transfusion. Admission diagnoses included: acute coronary syndrome , 42%; heart failure, 50%; cardiac arrest , 12%; and cardiogenic shock , 12%. Patients who received RBC transfusion had higher crude hospital mortality (19% vs. 8%, P<.001). RBC transfusion was associated with lower adjusted hospital mortality in patients with nadir Hgb <8 g/dL after propensity adjustment, including subgroups with acute coronary syndrome, cardiac arrest, or cardiogenic shock (all P <.01). RBC transfusion was not associated with lower adjusted hospital mortality in any subgroup of patients with nadir Hgb ≥8 g/dL. CONCLUSIONS: These observational data suggest the use of a Hgb threshold <8 g/dL for RBC transfusion in most CICU patients, although we could not exclude a potential benefit of RBC transfusion at a nadir Hgb of 8 to 9.9 g/dL; we did not observe any benefit from RBC transfusion at a nadir Hgb ≥10 g/dL.

Admission diagnosis and mortality risk prediction in a contemporary cardiac intensive care unit population.

BACKGROUND: Critical care risk scores can stratify mortality risk among cardiac intensive care unit (CICU) patients, yet risk score performance across common CICU admission diagnoses remains uncertain. METHODS: We evaluated performance of the Acute Physiology and Chronic Health Evaluation (APACHE)-III, APACHE-IV, Sequential Organ Failure Assessment (SOFA) and Oxford Acute Severity of Illness Score (OASIS) scores at the time of CICU admission in common CICU admission diagnoses. Using a database of 9,898 unique CICU patients admitted between 2007 and 2015, we compared the discrimination (c-statistic) and calibration (Hosmer-Lemeshow statistic) of each risk score in patients with selected admission diagnoses. RESULTS: Overall hospital mortality was 9.2%. The 3182 (32%) patients with a critical care diagnosis such as cardiac arrest, shock, respiratory failure, or sepsis accounted for >85% of all hospital deaths. Mortality discrimination by each risk score was comparable in each admission diagnosis (c-statistic 95% CI values were generally overlapping for all scores), although calibration was variable and best with APACHE-III. The c-statistic values for each score were 0.85-0.86 among patients with acute coronary syndromes, and 0.76-0.79 among patients with heart failure. Discrimination for each risk score was lower in patients with critical care diagnoses (c-statistic range 0.68-0.78) compared to non-critical cardiac diagnoses (c-statistic range 0.76-0.86). CONCLUSIONS: The tested risk scores demonstrated inconsistent performance for mortality risk stratification across admission diagnoses in this CICU population, emphasizing the need to develop improved tools for mortality risk prediction among critically-ill CICU patients.

Consensus modeling: Safer transfer learning for small health systems.

Predictive modeling is becoming an essential tool for clinical decision support, but health systems with smaller sample sizes may construct suboptimal or overly specific models. Models become over-specific when beside true physiological effects, they also incorporate potentially volatile site-specific artifacts. These artifacts can change suddenly and can render the model unsafe. To obtain safer models, health systems with inadequate sample sizes may adopt one of the following options. First, they can use a generic model, such as one purchased from a vendor, but often such a model is not sufficiently specific to the patient population and is thus suboptimal. Second, they can participate in a research network. Paradoxically though, sites with smaller datasets contribute correspondingly less to the joint model, again rendering the final model suboptimal. Lastly, they can use transfer learning, starting from a model trained on a large data set and updating this model to the local population. This strategy can also result in a model that is over-specific. In this paper we present the consensus modeling paradigm, which uses the help of a large site (source) to reach a consensus model at the small site (target). We evaluate the approach on predicting postoperative complications at two health systems with 9,044 and 38,045 patients (rare outcomes at about 1% positive rate), and conduct a simulation study to understand the performance of consensus modeling relative to the other three approaches as a function of the available training sample size at the target site. We found that consensus modeling exhibited the least over-specificity at either the source or target site and achieved the highest combined predictive performance.

Creating an atlas of normal tissue for pruning WSI patching through anomaly detection.

Patching whole slide images (WSIs) is an important task in computational pathology. While most of them are designed to classify or detect the presence of pathological lesions in a WSI, the confounding role and redundant nature of normal histology are generally overlooked. In this paper, we propose and validate the concept of an "atlas of normal tissue" solely using samples of WSIs obtained from normal biopsies. Such atlases can be employed to eliminate normal fragments of tissue samples and hence increase the representativeness of the remaining patches. We tested our proposed method by establishing a normal atlas using 107 normal skin WSIs and demonstrated how established search engines like Yottixel can be improved. We used 553 WSIs of cutaneous squamous cell carcinoma to demonstrate the advantage. We also validated our method applied to an external dataset of 451 breast WSIs. The number of selected WSI patches was reduced by 30% to 50% after utilizing the proposed normal atlas while maintaining the same indexing and search performance in leave-one-patient-out validation for both datasets. We show that the proposed concept of establishing and using a normal atlas shows promise for unsupervised selection of the most representative patches of the abnormal WSI patches.

Predictors of mortality by an artificial intelligence enhanced electrocardiogram model for cardiac amyloidosis.

AIMS: We aim to determine if our previously validated, diagnostic artificial intelligence (AI) electrocardiogram (ECG) model is prognostic for survival among patients with cardiac amyloidosis (CA). METHODS: A total of 2533 patients with CA (1834 with light chain amyloidosis (AL), 530 with wild-type transthyretin amyloid protein (ATTRwt) and 169 with hereditary transthyretin amyloid (ATTRv)] were included. An amyloid AI ECG (A2E) score was calculated for each patient reflecting the likelihood of CA. CA stage was calculated using the European modification of the Mayo 2004 criteria for AL and Mayo stage for transthyretin amyloid (ATTR). Risk of death was modelled using Cox proportional hazards, and Kaplan-Meier was used to estimate survival. RESULTS: Median age of the cohort was 67 [inter-quartile ratio (IQR) 59, 74], and 71.6% were male. The median overall survival for the cohort was 35.6 months [95% confidence interval (CI) 32.3, 39.5]. For AL, ATTRwt and ATTRv, respectively, median survival was 22.9 (95% CI 19.2, 28.2), 47.2 (95% CI 43.4, 52.3) and 61.4 (95% CI 48.7, 75.9) months. On univariate analysis, an increasing A2E score was associated with more than a two-fold risk of all-cause death. On multivariable analysis, the A2E score retained its importance with a risk ratio of 2.0 (95% CI 1.58, 2.55) in the AL group and 2.7 (95% CI 1.81, 4.24) in the ATTR group. CONCLUSIONS: Among patients with AL and ATTR amyloidosis, the A2E model helps to stratify risk of CA and adds another dimension of prognostication.

Artificial Intelligence of Arterial Doppler Waveforms to Predict Major Adverse Outcomes Among Patients Evaluated for Peripheral Artery Disease.

BACKGROUND: Patients with peripheral artery disease are at increased risk for major adverse cardiac events, major adverse limb events, and all-cause death. Developing tools capable of identifying those patients with peripheral artery disease at greatest risk for major adverse events is the first step for outcome prevention. This study aimed to determine whether computer-assisted analysis of a resting Doppler waveform using deep neural networks can accurately identify patients with peripheral artery disease at greatest risk for adverse outcome events. METHODS AND RESULTS: Consecutive patients (April 1, 2015, to December 31, 2020) undergoing ankle-brachial index testing were included. Patients were randomly allocated to training, validation, and testing subsets (60%/20%/20%). Deep neural networks were trained on resting posterior tibial arterial Doppler waveforms to predict major adverse cardiac events, major adverse limb events, and all-cause death at 5 years. Patients were then analyzed in groups based on the quartiles of each prediction score in the training set. Among 11 384 total patients, 10 437 patients met study inclusion criteria (mean age, 65.8±14.8 years; 40.6% women). The test subset included 2084 patients. During 5 years of follow-up, there were 447 deaths, 585 major adverse cardiac events, and 161 MALE events. After adjusting for age, sex, and Charlson comorbidity index, deep neural network analysis of the posterior tibial artery waveform provided independent prediction of death (hazard ratio [HR], 2.44 [95% CI, 1.78-3.34]), major adverse cardiac events (HR, 1.97 [95% CI, 1.49-2.61]), and major adverse limb events (HR, 11.03 [95% CI, 5.43-22.39]) at 5 years. CONCLUSIONS: An artificial intelligence-enabled analysis of Doppler arterial waveforms enables identification of major adverse outcomes among patients with peripheral artery disease, which may promote early adoption and adherence of risk factor modification.

Differentiating clinical characteristics of perianal inflammatory bowel disease from perianal hidradenitis suppurativa.

BACKGROUND: Perianal draining tunnels in hidradenitis suppurativa (HS) and perianal fistulizing inflammatory bowel disease (IBD) present diagnostic and management dilemmas. METHODS: We conducted a retrospective chart review of patients with perianal disease evaluated at Mayo Clinic from January 1, 1998, through July 31, 2021. Patients' demographic and clinical data were extracted, and 28 clinical features were collected. After experimenting with several machine learning techniques, random forests were used to select the 15 most important clinical features to construct the diagnostic prediction model to distinguish perianal HS from fistulizing perianal IBD. RESULTS: A total of 263 patients were included (98 with HS, 100 with IBD, and 65 with both IBD and HS). Patients with HS had a higher mean body mass index, a higher smoking rate, and more commonly showed cutaneous manifestations of tunnels and comedones, while fistulas, abscesses, induration, anal tags, ulcers, and anal fissures were more common in patients with IBD. In addition to having lesions in the perianal area, patients with IBD often had lesions in the buttocks and perineum, while those with HS had additional lesions in the axillae and groin. Among the statistically significant features, the 15 most important were identified by random forest: fistula, tunnel, digestive symptom, knife-cut ulcer, perineum, body mass index, age, axilla, abscess, tags, smoking, groin, genital cutaneous edema, erythema, and bilateral/unilateral. CONCLUSIONS: The results of this study may help differentiate perianal lesions, especially perineal HS and fistulizing perineal IBD, and provide promise for a better therapeutic outcome.

Development of the lichen planus quality of life questionnaire (LPQoL) informed by expert clinician input and patient feedback: a retrospective survey study.

Lichen planus (LP) can affect multiple body sites including skin, mucosae, scalp and nails, causing considerable impact on patients' quality of life. Currently, there are no LP patient-reported outcome measures (PROMs) that address all body sites potentially affected by LP. We developed a LP Quality of Life Questionnaire (LPQoL), informed by an expert consortium and patient survey study, to address this gap. The study was approved by our institution's Institutional Review Board. First, a 22-item LPQoL was designed with input from LP experts at our institution. The tool was then optimized by garnering input from patients recently diagnosed with LP, who were asked to complete the LPQoL, as well as the Dermatology Life Quality Index (DLQI) and a feedback form about the LPQoL. Fifty-eight of 150 patients (39% response rate) returned the questionnaire. Mean DLQI score was 4.9 ± 5.6 SD (range 0-25) and mean LPQoL score was 13.6 ± 10.4 SD (range 0-54). LPQoL score was positively correlated with DLQI score (r = 0.79; p < 0.001). Forty-nine out of 56 (88%) and 6/56 (11%) rated the LPQoL as 'very easy' or 'fairly easy' to complete, respectively. Based on participants' feedback, we increased the recall period from one week to one month and added questions on esophageal involvement. With iterative input from LP experts and patients, we developed a LPQoL to address the gap in a multi-site PROM specific to LP. This is a pilot study and there is ongoing validation studies; therefore, this measure should not be used in clinical practice or research until validated.

Postdevelopment Performance and Validation of the Artificial Intelligence-Enhanced Electrocardiogram for Detection of Cardiac Amyloidosis.

BACKGROUND: We have previously applied artificial intelligence (AI) to an electrocardiogram (ECG) to detect cardiac amyloidosis (CA). OBJECTIVES: In this validation study, the authors observe the postdevelopment performance of the AI-enhanced ECG to detect CA with respect to multiple potential confounders. METHODS: Amyloid patients diagnosed after algorithm development (June 2019-January 2022) with a 12-lead ECG were identified (n = 440) and were required to have CA. A 15:1 age- and sex-matched control group was identified (n = 6,600). Area under the receiver operating characteristic (AUC) was determined for the cohort and subgroups. RESULTS: The average age was 70.4 ± 10.3 years, 25.0% were female, and most patients were White (91.3%). In this validation, the AI-ECG for amyloidosis had an AUC of 0.84 (95% CI: 0.82-0.86) for the overall cohort and between amyloid subtypes, which is a slight decrease from the original study (AUC 0.91). White, Black, and patients of "other" races had similar algorithm performance (AUC >0.81) with a decreased performance for Hispanic patients (AUC 0.66). Algorithm performance shift over time was not observed. Low ECG voltage and infarct pattern exhibited high AUC (>0.90), while left ventricular hypertrophy and left bundle branch block demonstrated lesser performance (AUC 0.75 and 0.76, respectively). CONCLUSIONS: The AI-ECG for the detection of CA maintained an overall strong performance with respect to patient age, sex, race, and amyloid subtype. Lower performance was noted in left bundle branch block, left ventricular hypertrophy, and ethnically diverse populations emphasizing the need for subgroup-specific validation efforts.

Temporal Trends, Predictors and Outcomes of Inpatient Palliative Care Use in Cardiac Arrest Complicating Acute Myocardial Infarction.

BACKGROUND: Utilization of inpatient palliative care services (PCS) has been infrequently studied in patients with cardiac arrest complicating acute myocardial infarction (AMI-CA). METHODS: Adult AMI-CA admissions were identified from the National Inpatient Sample (2000-2017). Outcomes of interest included temporal trends and predictors of PCS use and in-hospital mortality, length of stay, hospitalization costs and discharge disposition in AMI-CA admissions with and without PCS use. Multivariable logistic regression and propensity matching were used to adjust for confounding. RESULTS: Among 584,263 AMI-CA admissions, 26,919 (4.6%) received inpatient PCS. From 2000 to 2017 PCS use increased from <1% to 11.5%. AMI-CA admissions that received PCS were on average older, had greater comorbidity, higher rates of cardiogenic shock, acute organ failure, lower rates of coronary angiography (48.6% vs 63.3%), percutaneous coronary intervention (37.4% vs 46.9%), and coronary artery bypass grafting (all p < 0.001). Older age, greater comorbidity burden and acute non-cardiac organ failure were predictive of PCS use. In-hospital mortality was significantly higher in the PCS cohort (multivariable logistic regression: 84.6% vs 42.9%, adjusted odds ratio 3.62 [95% CI 3.48-3.76]; propensity-matched analysis: 84.7% vs. 66.2%, p < 0.001). The PCS cohort received a do- not-resuscitate status more often (47.6% vs. 3.7%), had shorter hospital stays (4 vs 5 days), and were discharged more frequently to skilled nursing facilities (73.6% vs. 20.4%); all p < 0.001. These results were consistent in the propensity-matched analysis. CONCLUSIONS: Despite an increase in PCS use in AMI-CA, it remains significantly underutilized highlighting the role for further integrating of these specialists in AMI-CA care.

Estimation of manual wheelchair-based activities in the free-living environment using a neural network model with inertial body-worn sensors.

Shoulder pain is common in manual wheelchair (MWC) users. Overuse is thought to be a major cause, but little is known about exposure to activities of daily living (ADLs). The study goal was to develop a method to estimate three conditions in the field: (1) non-propulsion activity, (2) MWC propulsion, and (3) static time using an inertial measurement unit (IMU). Upper arm IMU data were collected as ten MWC users performed lab-based MWC-related ADLs. A neural network model was developed to classify data as non-propulsion activity, propulsion, or static, and validated for the lab-based data collection by video comparison. Six of the participants' free-living IMU data were collected and the lab-based model was applied to estimate daily non-propulsion activity, propulsion, and static time. The neural network model yielded lab-based validity measures ≥0.87 for differentiating non-propulsion activity, propulsion, and static time. A quasi-validation of one participant's field-based data yielded validity measures ≥0.66 for identifying propulsion. Participants' estimated mean daily non-propulsion activity, propulsion, and static time ranged from 158 to 409, 13 to 25, and 367 to 609 min, respectively. The preliminary results suggest the model may be able to accurately identify MWC users' field-based activities. The inclusion of field-based IMU data in the model could further improve field-based classification.