Despite growing interest from both patients and healthcare providers, there is little clinical guidance on how mobile apps should be utilized to add value to patient care. We categorize apps according to their functionality (e.g. preventative behavior change, digital self-management of a specific condition, diagnostic) and discuss evidence for effectiveness from published systematic reviews and meta-analyses and the relevance to patient care. We discuss the limitations of the current literature describing clinical outcomes from mHealth apps, what FDA clearance means now (510(k)/de novo FDA clearance) and in the future. We discuss data security and privacy as a major concern for patients when using mHealth apps. Patients are often not involved in the development of mobile health guidelines, and professionals’ views regarding high-quality health apps may not reflect patients’ views. We discuss efforts to develop guidelines for the development of safe and effective mHealth apps in the US and elsewhere and the role of independent app reviews sites in identifying mHealth apps for patient care. There are only a small number of clinical scenarios where published evidence suggests that mHealth apps may improve patient outcomes.

Do-it-yourself automated insulin delivery systems for people living with type 1 diabetes use commercially available continuous glucose sensors and insulin pumps linked by unregulated open source software. Uptake of these systems is increasing, with growing evidence suggesting that positive glucose outcomes may be feasible. Increasing interest from people living with, or affected by, type 1 diabetes presents challenges to healthcare professionals, device manufacturers and regulators as the legal, governance and risk frameworks for such devices are not defined. We discuss the data, education, policy, technology and medicolegal obstacles to wider implementation of DIY systems and outline the next steps required for a co-ordinated approach to reducing variation in access to a technology that has potential to enable glucose self-management closer to target.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Fueled by breakthrough technology developments, the biological, biomedical, and behavioral sciences are now collecting more data than ever before. There is a critical need for time- and cost-efficient strategies to analyze and interpret these data to advance human health. The recent rise of machine learning as a powerful technique to integrate multimodality, multifidelity data, and reveal correlations between intertwined phenomena presents a special opportunity in this regard. However, machine learning alone ignores the fundamental laws of physics and can result in ill-posed problems or non-physical solutions. Multiscale modeling is a successful strategy to integrate multiscale, multiphysics data and uncover mechanisms that explain the emergence of function. However, multiscale modeling alone often fails to efficiently combine large datasets from different sources and different levels of resolution. Here we demonstrate that machine learning and multiscale modeling can naturally complement each other to create robust predictive models that integrate the underlying physics to manage ill-posed problems and explore massive design spaces. We review the current literature, highlight applications and opportunities, address open questions, and discuss potential challenges and limitations in four overarching topical areas: ordinary differential equations, partial differential equations, data-driven approaches, and theory-driven approaches. Towards these goals, we leverage expertise in applied mathematics, computer science, computational biology, biophysics, biomechanics, engineering mechanics, experimentation, and medicine. Our multidisciplinary perspective suggests that integrating machine learning and multiscale modeling can provide new insights into disease mechanisms, help identify new targets and treatment strategies, and inform decision making for the benefit of human health.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Patient online health searching is now commonplace, however, the accuracy of patient generated differentials for new symptoms and potential for patient anxiety are concerns. We aimed primarily to determine the accuracy of patient generated differentials for new symptoms with and without online searching, and secondarily, to evaluate the impact of searching on anxiety levels. In the waiting room prior to seeing a clinician, 300 patients with new symptoms were randomly assigned 1:1:1 to Google searching with health related features including a symptom search tool vs Google searching with health related features disabled vs no searching. Participants were 18 years or older and presenting to the emergency department of an urban academic medical center with new low-acuity symptoms that were not due to exacerbation of a chronic condition. Search groups received access on a tablet/smartphone to Google searching with or without health related features. Both search groups could access any websites; health related features led the patient to common diagnoses and physician-validated information. The primary outcome was accuracy of the patient generated differential assessed by matching at least two of the top three diagnoses on the clinician’s differential. A secondary outcome was anxiety by a visual analogue scale. Patients were a median of 33.1 (IQI 26.2–45.9) years old, 60% women, 63% black, 82% had a high school education or less, and 45.7% reported having performed an online search prior to presentation. Search group patients spent a median of 3.82 (2.53–5.72) minutes searching online. Similar proportions of patients in each group matched at least two of three clinician diagnoses: 27.0% and 28.3% for Google searching with and without health related features vs 23.8% in the no search group. Patients in the search groups had a similar odds of matching ≥2/3 diagnoses as the no search group [OR (95% CI): 1.23 (0.70–2.13), p = 0.47]. Anxiety was unchanged with online searching. In conclusion, brief online searching in the waiting room did not improve accuracy of patient generated differential diagnoses for new symptoms. The absence of an increase in patient anxiety provides reassurance for subsequent work to refine and investigate online symptom search tools.

When mobile health (mHealth) applications (apps) are investigated, the question of the proper control condition arises. Normally, the randomized controlled trial (RCT) is seen as the gold standard when testing efficacy of clinical interventions. Yet, mHealth apps rarely comprise innovative treatments but rather provide established treatments digitally. The classical RCT utilizing a placebo or waiting group condition may not always be the suitable methodology, since non-treatment is not appropriate if a disease urges treatment and the development of chronic disease needs to be prevented. The present commentary discusses conceivable control conditions in mHealth trials and illustrates their limitations.

End-stage liver disease (ESLD) is associated with cognitive impairment ranging from subtle alterations in attention to overt hepatic encephalopathy that resolves after transplant. Natural language processing (NLP) may provide a useful method to assess cognitive status in this population. We identified 81 liver transplant recipients with ESLD (4/2013–2/2018) who sent at least one patient-to-provider electronic message pre-transplant and post-transplant, and matched them 1:1 to “healthy” controls—who had similar disease, but had not been evaluated for liver transplant—by age, gender, race/ethnicity, and liver disease. Messages written by patients pre-transplant and post-transplant and controls was compared across 19 NLP measures using paired Wilcoxon signed-rank tests. While there was no difference overall in word length, patients with Model for End-Stage Liver Disease Score (MELD) ≥ 30 (n = 31) had decreased word length in pre-transplant messages (3.95 [interquartile range (IQR) 3.79, 4.14]) compared to post-transplant (4.13 [3.96, 4.28], p = 0.01) and controls (4.2 [4.0, 4.4], p = 0.01); there was no difference between post-transplant and controls (p = 0.4). Patients with MELD ≥ 30 had fewer 6+ letter words in pre-transplant messages (19.5% [16.4, 25.9] compared to post-transplant (23.4% [20.0, 26.7] p = 0.02) and controls (25.0% [19.2, 29.4]; p = 0.01). Overall, patients had increased sentence length pre-transplant (12.0 [9.8, 13.7]) compared to post-transplant (11.0 [9.2, 13.3]; p = 0.046); the same was seen for MELD ≥ 30 (12.3 [9.8, 13.7] pre-transplant vs. 10.8 [9.6, 13.0] post-transplant; p = 0.050). Application of NLP to patient-generated messages identified language differences—longer sentences with shorter words—that resolved after transplant. NLP may provide opportunities to detect cognitive impairment in ESLD.

One broad goal of biomedical informatics is to generate fully-synthetic, faithfully representative electronic health records (EHRs) to facilitate data sharing between healthcare providers and researchers and promote methodological research. A variety of methods existing for generating synthetic EHRs, but they are not capable of generating unstructured text, like emergency department (ED) chief complaints, history of present illness, or progress notes. Here, we use the encoder-decoder model, a deep learning algorithm that features in many contemporary machine translation systems, to generate synthetic chief complaints from discrete variables in EHRs, like age group, gender, and discharge diagnosis. After being trained end-to-end on authentic records, the model can generate realistic chief complaint text that appears to preserve the epidemiological information encoded in the original record-sentence pairs. As a side effect of the model's optimization goal, these synthetic chief complaints are also free of relatively uncommon abbreviation and misspellings, and they include none of the personally identifiable information (PII) that was in the training data, suggesting that this model may be used to support the de-identification of text in EHRs. When combined with algorithms like generative adversarial networks (GANs), our model could be used to generate fully-synthetic EHRs, allowing healthcare providers to share faithful representations of multimodal medical data without compromising patient privacy. This is an important advance that we hope will facilitate the development of machine-learning methods for clinical decision support, disease surveillance, and other data-hungry applications in biomedical informatics.