Precision diagnosis: a view of the clinical decision support systems (CDSS) landscape through the lens of critical care

Abstract

Improving diagnosis and treatment depends on clinical monitoring and computing. Clinical decision support systems (CDSS) have been in existence for over 50 years. While the literature points to positive impacts on quality and patient safety, outcomes, and the avoidance of medical errors, technical and regulatory challenges continue to retard their rate of integration into clinical care processes and thus delay the refinement of diagnoses towards personalized care. We conducted a systematic review of pertinent articles in the MEDLINE, US Department of Health and Human Services, Agency for Health Research and Quality, and US Food and Drug Administration databases, using a Boolean approach to combine terms germane to the discussion (clinical decision support, tools, systems, critical care, trauma, outcome, cost savings, NSQIP, APACHE, SOFA, ICU, and diagnostics). References were selected on the basis of both temporal and thematic relevance, and subsequently aggregated around four distinct themes: the uses of CDSS in the critical and surgical care settings, clinical insertion challenges, utilization leading to cost-savings, and regulatory concerns. Precision diagnosis is the accurate and timely explanation of each patient’s health problem and further requires communication of that explanation to patients and surrogate decision-makers. Both accuracy and timeliness are essential to critical care, yet computed decision support systems (CDSS) are scarce. The limitation arises from the technical complexity associated with integrating and filtering large data sets from diverse sources. Provider mistrust and resistance coupled with the absence of clear guidance from regulatory bodies further retard acceptance of CDSS. While challenges to develop and deploy CDSS are substantial, the clinical, quality, and economic impacts warrant the effort, especially in disciplines requiring complex decision-making, such as critical and surgical care. Improving diagnosis in health care requires accumulation, validation and transformation of data into actionable information. The aggregate of those processes—CDSS—is currently primitive. Despite technical and regulatory challenges, the apparent clinical and economic utilities of CDSS must lead to greater engagement. These tools play the key role in realizing the vision of a more ‘personalized medicine’, one characterized by individualized precision diagnosis rather than population-based risk-stratification.

Improving diagnosis in health care

On September 22, 2015, the Institute of Medicine (IOM) of the National Academy of Sciences released the third report in its Quality Chasm series. This report, which is titled Improving Diagnosis in Health Care, presents a sobering look at the current rates of diagnostic error across the spectrum of clinical practice [1]. The reported rates suggest that diagnostic errors are common and substantially affect the health and well-being of patients. The IOM report defines diagnostic error consequent to one of two failures: a failure to establish an accurate and timely explanation of the patient’s health problem(s); or, a failure to communicate that explanation to the patient. Nowhere is diagnostic accuracy and timeliness more important than in the surgical and critical care settings, where making the correct decision so strongly affects morbidity and mortality. This diagnostic accuracy is founded on monitoring and computing that drives clinical decision support. Accordingly, we review the history and assessed the current state of clinical decision support systems (CDSS) in these surgical and critical care settings.

Combining terms germane to the discussion (clinical decision support, tools, systems, trauma, acute, outcome, cost, NSQIP, APACHE, ICU, and diagnostics), the authors queried the MEDLINE/PubMed database, relying heavily on a few systematic reviews to further guide their search. Results were cross-referenced and the authors eliminated case reports, editorials, and references that contained duplicate information. Special care was given to selecting references focused on the acute care space. Additional search engines (Institute of Medicine, RAND Health, Agency for Health Care Research and Quality, Food and Drug Administration) were targeted to pull reports focused on CDSS. Of those combined resources, 91 were eventually selected on the basis of both temporal and thematic relevance (i.e. critical care, when available) (Table 1).

Table 1 Search strategy

A brief history of clinical decision support systems

Clinical decision support systems (CDSS) were proposed more than a half-century ago. In 1959, Ledley and Lusted described logic and probability in medical decision-making [2]. This seminal work, which focused on the use of computers to improve medical diagnosis, catalyzed efforts to address gaps in clinical delivery failures [3–9]. The contributions of Sheppard and Kouchoukos [10] and Sheppard et al. [11] on the use of automated measurements and rules as a potential substitute for judgment in the critical care space further advanced the field [10, 11]. Yet it took another thirty-two years for a consensus definition of what constitutes a clinical decision support system (CDSS) to emerge: ‘active knowledge systems that use two or more items of patient data to generate case-specific advice’ [12].

In the wake of the prior two influential Institute of Medicine reports entitled To Err is Human [13] and Crossing the Quality Chasm [14], each of which pointed to faulty systems and processes as the root cause of human medical error, the healthcare industry witnessed a steady rise in the use of CDSS. The use of these medical decision-making tools touches a range of disciplines across the healthcare landscape, from dental [15] to imaging [16]. CDSS have been reported to favorably influence quality and patient safety [17–20], to promote prevention and optimal treatment [21–23], to reduce medical errors [24, 25], and to improve outcomes [26, 27].

A closer look at these studies reveals that the primary emphasis of CDSS has been on “what to do?” more than “is the diagnosis accurate?”. A 2005 study by Garg et al. offers a thorough analysis of electronic CDSS on provider performance and patient outcomes as part of a systematic review of one hundred studies categorized as systems for diagnosis (10 trials), reminder systems for prevention (21 trials), systems for disease management (40 studies), and systems for drug dosing and drug prescribing (29 trials). In this systematic review, the authors reported on increased practitioner performance (e.g. adherence to guidelines, recognition of functional status problems, time to order treatment) in 62 of the 97 studies selected (64 %), where improvement was defined as a statistically significant positive effect on at least 50 % of outcomes measured [28]. Such improvement only hints at the potential of CDSS to close performance gaps in critical care. For example, failure to discern a common critical state, the adult respiratory distress syndrome (ARDS), is associated with failure to consistently provide evidence-based low-tidal volume ventilation in 13.4 to 45 % of cases [29–31].

All of this begs two questions: “Are computers helping physicians reach the correct diagnosis?”; and “Does having a correct diagnosis materially improve treatment and outcome?”

Current management of complex disease and critical illness remains largely dependent upon accumulated experience and pattern repetition. Such reliance predisposes to various biases, with predictable results: under treatment of some patients, who subsequently have high complication rates, and the overtreatment of others, leading to prolonged pain, the potential for perioperative complications and the expenditure of valuable healthcare resources. A vivid example is found in the management of open wounds, where dehiscence (the failure of a wound to stay closed after the wound edges are approximated, a skin graft is placed or the wound is covered by other tissue) leads to delays in definitive closure, marked increase in costs, worse pain and nutritional setbacks. It is therefore important to attempt closure only when success is highly likely. Yet reliance on accumulated experience and pattern recognition results in failure rates of 15 % [32]. CDSS, if properly implemented, have the potential to not only contain costs but also improve outcomes by ensuring the optimal treatment is delivered at the right time.

Indeed, the imperative to deliver the right care, right now, every time is central to the work of our research consortium, the surgical critical care initiative (www.sc2i.org) [33] and will resonate with related disciplines requiring complex decision-making. Errors in judgment, in decisions, and in performance lead to protracted lengths of stay [34], substantial increases in hospital costs [35], and considerable escalations in both perioperative and late mortality rates [36]. We believe CDSS, used in conjunction with reproducible treatment protocols, can play a pivotal role in mitigating these adverse clinical events. Indeed, the work of Morris et al. [37] raises the question of replicable methods to reduce the error rate amongst clinical decision makers.

Clinical decision support in the critical and surgical care settings

The purpose of CDSS in the critical care space, whether rule-based or probabilistic, is to augment, rather than replace, a provider’s ability to infer a correct diagnosis. This augmented inference is different from rule-based negative feedback control models that have been shown effective in fully closed-loops titrating specific therapies to isolated endpoints (such as targeting blood pressure) [38, 39]. Critical illness and injury—particularly that separate from perioperative trajectories following elective surgery—have sufficiently complex dynamics to justify maintaining humans in the control loop. CDSS validated through controlled clinical trials can nevertheless frame options based on evidence. This is foundational to “Right Care” (i.e. accurate diagnosis and treatment plan). Moreover, by presenting those options and probabilistic outcomes, CDSS can accelerate the implementation of the selected tests or treatments. This is foundational to “Right Now”.

Current structure promotes filtering and integrating patient-specific discrete blocks of data to reduce information overload as a strategy to improve medical decision-making. The architecture of these tools is standardized in that they typically combine three components into a decision support platform: the clinical and/or physiological information, an underlying model that analyzes the data, and an interface to collect inputs and display outputs. CDSS organized in this fashion can support a wide range of tasks [40]. Of particular interest to the critical and surgical care settings are tasks pertaining to diagnosis (matching a patient’s signs and symptoms) and treatment planning (operationalizing guidelines following specific diagnoses). We will return to the issue of diagnostic precision and guideline specificity later in this review.

The overall influence of CDSS on clinical practice, whether it touches on quality and patient safety, clinical outcomes, or the avoidance of medical errors, is well documented [41, 42]. In cases where large, variable, and multi-source data are being processed, CDSS affect the quality of the decisions being made, leading to improved clinical outcomes. We believe this is particularly true in instances where multi-disciplinary teams care under strained workload and time pressures [43], such as trauma or complex surgical care (Table 2). In 1998, Lanceley et al. highlighted the struggle to balance provider input (surgical, medical, and diagnostic expertise) with health-care guidelines to achieve superior outcomes.

Table 2 Non-exhaustive list of CDSS used in surgical and intensive care settings

CDSS have historically been divided into knowledge-driven (e.g., powered by if-then statements) and data-driven (machine learning from large datasets implementing Bayesian or neural networks, fuzzy logic theories, symbolic reasoning) support systems. While all CDSS fundamentally attempt to address a four-prong problem (i.e. processing an increasingly large and heterogeneous volume of medical information, accounting for missing clinical or physiologic information, physician overload, and medical errors), their complexity can vary substantially from tool to tool [44]. Whereas alert systems may follow strict IF-THEN logic rules to report on an adverse drug interaction, systems powered by artificial intelligence can calculate the probability of an emerging adverse condition, even in the presence of missing data. We think this is where methods such as fuzzy logic, by virtue of their connectives and inference rules, are better suited to the critical-care space, a discipline characterized by complex causal chains and uncertainty [45]. In the current era of advanced computing that allows for near instantaneous identification and analysis of reference populations, the distinction between knowledge-driven and data-driven approaches appears to have been blurred. Our systematic review of the literature yielded predominantly data-driven support systems since they have been most usefully applied for cases requiring complex medical and surgical decision-making [46].

Although first introduced in 1984 [47], CDSS for critical care are exclusively knowledge-based products that, while generally effective, have been traditionally driven by clinical and physiological inputs; biomarker data, when used, are limited to the most conventional laboratory values. One example of a widely used clinical decision support tool is the outcomes-driven National Surgical Quality Improvement Program (NSQIP) risk calculator, developed by the Veterans Health Administration and implemented in the private sector by the American College of Surgeons [48–50]. This CDSS, using data from 400 hospitals and 1.4 million patients, estimates patient-specific postoperative complication risks for more than 1500 unique surgical procedures across multiple specialties. The effectiveness of the NSQIP Risk Calculator in containing surgical complications is the subject of rigorous, on-going, monitoring and reporting. A 2007 study released by the American College of Surgeons asserts that use of this particular CDSS was associated with a 47 % drop in postoperative mortality and a 43 % drop in postoperative morbidity rates from 1991 to 2006 [51]. The success of the NSQIP Risk Calculator led to further development of predictive tools, to include one which prospectively identifies patients at high-risk for surgical complications based on their own characteristics [52]. Other tools, like the Physiological and Operative Severity Score for the enumeration of mortality and morbidity (POSSUM) have demonstrated their utility in limiting post-operative complications [53]. Surgical wound infections are the second most common nosocomial disease, occurring in 2 to 5 % of all surgeries and 20 % of abdominal surgeries [54]; a simple automated reminder to the anesthesia team and the surgeon to administer prophylactic antibiotics can favorably influence surgical outcomes [55].

Two other tools used predominantly for care quality comparisons, performance improvements and research in the critical care setting are the acute physiology and chronic health evaluation (APACHE), introduced in 1985 to rate the severity of illness for adult patients admitted to intensive care units, and the sequential organ failure assessment (SOFA), which also provides a valuable prognostic estimate of in-hospital survival [56]. The predictive values generated by the underlying models rely on a variety of physiologic and laboratory inputs (demographics, vitals, oxygenation, chemistry, hematology). In 2014, Gultepe et al. [57] reported further improvements to these approaches, using advanced machine-learning techniques to develop a model capable of identifying patients at high risk for hyperlactatemia. While the aforementioned tools prove helpful in predicting mortality rates for groups of critically ill patients [58], we concur with the widely recognized point that, in light of their limitations, none can be a perfect substitute for the provider’s expert clinical judgment in the care of an individual patient [59]. Therein lies the limitation of ‘current-gen’ tools, which tend to achieve optimal group outcomes through risk stratification rather than generate personalized assessment and risk predictions; yet this latter path towards precision diagnosis and recommendation is precisely where ‘next-gen’ clinical decision support tools should go, as suggested by Loghmanpour et al. [60] in their 2014 assessment of a Bayesian-based model to treat end-stage heart failure and reinforced by the release of the newly released, and aforementioned, ‘Improving Diagnosis in Health Care’ IOM report.

The performance of CDSS in emergency departments seems variable. While one study may question the accuracy of the tool as an arbiter in diagnostic decision support [61], another points to the robustness of an artificial neural network as a viable first-screening methodology to differentiate cardiac from non-cardiac chest pain [62]. Differences in underlying methodologies notwithstanding, and in the context of an increased development of CDSS in the critical care arena, more research in this field appears warranted to address the utility—and physician acceptance—of these decision-support tools.

Although precision medicine is not a new concept, substantial advances in both genotyping and microarray/biochips in the past two decades [63] elevate the potential of tailoring medical treatment to a patient’s biological characteristics, needs, and preference during all stages of care. Our ability to leverage ‘big data’ [64] should augment the arsenal of decision-support models currently in-use in the operating room (e.g. anesthesia information management systems—AIMS) or Intensive Care Units (e.g. Architecture for intensive care unit decision support—ACUDES; Rhea). The AIMS and ACUDES, like most CDSS, are representative of current generation of electronic health records, capturing discrete packets of information in relational databases and processing these inputs, for intraoperative quality assurance and research in the case of the former and to manage a patient’s temporal evolution in the case of the latter.

A consequence of this explosion of data, and of particular interest to the critical and surgical care settings, is the implementation of machine learning CDSS that are able to suggest biomarker causal dependencies, even in the absence of data sets. This is a relatively new field, one in which promising results have already been reported [65]; in their five-year study, Forsberg et al. estimated the survival of 189 consecutive patients with operable skeletal metastases by integrating multifaceted clinical data into a prognostic model. The underlying statistical method used for this analysis, Bayesian Belief Network (BBN), yielded a robust and accurate CDSS.

Challenges to development and clinical insertion

While a few risk-stratification CDSS have proven durable in the critical-care setting, further development and subsequent deployment of patient-specific predictive tools has not occurred. The reasons are many: converting data into a clinically relevant model, difficulty in integrating the tool into the clinical workflow (to include Electronic Health Records), and ethical/legal implications of ignoring or overriding recommendations.

The literature also points to providers resisting utilization, yet another obstacle to the widespread implementation of CDSS [66–68]. Several factors fuel this challenge: the system not obscuring the physician’s autonomy and responsibility; the availability of hardware, technical support and training; the integration of the system into workflows; and the relevance and timeliness of the recommendations provided. Improvements in clinical utilization are largely dependent on these dynamics yet bump against a natural inclination to resist change (e.g. threatening independence of thought, frustrating the belief one is already practicing in an evidence-based fashion). Barriers to adoption extend beyond the clinical realm to include wider organizational concerns of altering conservative behavior across institutional functions (e.g. clinical, research, administrative). Strategies pioneered by Morris and colleagues towards the development and distribution of replicable computer protocols have met with some success [69, 70]. In a study focused on multi-site ventilator management, East et al. also reported on the benefits of implementing a CDSS to improve patient morbidity [71]. These results emphasize that a unity of vision across disciplines, within (and potentially across) organizations is prerequisite to the successful implementation of decision-support systems. Responding to the aforementioned challenges is paramount if CDSS are to successfully support improvements in clinical outcomes, safety, and efficiency.

Yet the development of reliable tools appears to be the single largest hurdle to overcome. The availability of electronic patient data, fragmented across multiple hospital databases and systems (laboratory, microbiology, pathology, vital signs, radiology, surgery, etc.), remains a challenge, one which is of particular import for machine learning applications [72, 73]. While the challenge extends to any developed system attempting to achieve interoperability with established EHRs, it is of specific relevance to those disciplines that require the intelligent filtering of massive amounts of data, at a time when both precision diagnosis and optimal method of treatment are vital (e.g. critical and surgical care).

The goal of integrating and extracting useful or new electronic data across heterogeneous repositories remains elusive [74], but the need for CDSS to both manage resource utilization more efficiently and improve the quality of care delivered are well understood. However, their limited availability impedes the advancement of best practices [75] and represent a missed opportunity to reduce undesirable variation in care [69] and improve outcomes [76].

Reports on CDSS adoption-rates, and the adoption rates themselves, in critical care settings are somewhat limited. One informative study focuses on the decision support for safer surgery (DS3) tool. Results of a survey conducted with 23 surgeons on 1006 patients point to attitudes ranging from ‘neutral’ to ‘slightly positive’, with results encouraging adoption rate correlated to a patient’s higher estimated risk for mortality and morbidity [77]. An additional study suggests that reactions to CDSS may differ based on the perceived intent of the tool, whether it reminds providers of things they intend to do, provides information when clinicians are unsure what to do, corrects errors, or recommends an adjustment to the treatment plan [5]. Indeed, acceptance and adoption is likely to be prolonged and sequential, much the way drivers responded to GPS-enabled maps with improved computer-enabled situation awareness leading to trust in, and adoption of, the computed route recommendations.

Consideration must also be given to the ongoing applicability of the CDSS. Their relevance in the clinical sphere relies on their ability to perform optimally throughout their lifecycle [78], which requires some type of feedback loop to ensure dynamic models’ continuous testing and upgrading, as needed. These sustainment activities, which may require the establishment of distinct organizational units [79], carry with them a commitment of resources to ensure CDSS are effective in the long-term.

Reports on the costs of developing and implementing a CDSS are scarce, perhaps because these efforts are predominantly supported via research funding. A 2008 study does provide useful insight as it pertains to the deployment of a decision support tool providing prescribers with patient-specific maximum dosing recommendations based on renal function. Although limited in scope (94 alerts for 62 drugs), this project took no fewer than 924.5 man-hours and $49,669 to reach maturity [80], lending weight to the argument that an extensive commitment of resources, to include clinical staff, is required for the development, testing, and deployment of a CDSS. There is evidence that both the lack of funding and the absence of standardization across homegrown systems have thus far hindered the widespread development and deployment of CDSS [26, 81].

The advent of mobile platforms (e.g. notebooks, tablets, phones) may introduce an opportunity to facilitate the deployment of ‘connected-anytime-anywhere’ clinical decision support tools. However, the successful insertion of these tools in the clinical environment will likely depend on their ability to effectively and securely communicate with EHRs to either pull (e.g. value extraction to inform predictive models) or push (e.g. report on outcomes) data. There appears to be fertile ground for these points-of-care devices to blossom, provided they are fully integrated into the existing clinical IT infrastructure.

CDSS utilization can lead to substantial cost-savings

In 2011, inpatient care accounted for 31.5 % of the USA’s $2.7 trillion health expenditures; the same year, annual spending by hospitals grew 4.3 %, out-pacing overall health payments by 0.4 % [82]. A limited number of manuscripts report on the question of cost as an outcome [83] and yet the question becomes increasingly relevant under a ‘managed care’ framework with limited resources and funding. Available data, however sparse, suggest that investments in CDSS technologies that advance the principles of ‘The Right Care, Right Now; for Every Patient, Every Time’ are likely to have large favorable returns on those investments.

A review of 22 randomized controlled trials (CDSS implemented in real-clinical settings and used by health professionals to aid decision-making at the point of care) indicated a decrease in the number of interventions with a resulting reduction in costs when compared to control groups [84]. A similar cross-sectional analysis of urban hospitals in Texas concluded that higher use of decision-support tools was associated with a 16 % decrease in complications and $538 savings per encounter across all hospital admissions [18].

One of the more enlightening of the cost-focused studies, itself grounded in data generated from 1200 major surgery patients, estimates the average cost of cases with complications leading to a patient’s death (Grade 4) to be $159,345 [85], compared to $27,946 for uneventful cases. A more recent study estimates that reduced surgical complications (site infection, renal failure, graft/prosthesis, wound disruption) can add $2.2 million per 10,000 cases to hospital costs [86].

The impact of CDSS utilization on healthcare economics can further be gleaned from two randomized controlled trials assessing the impact of computed CDSS on antimicrobial utilization; these studies reported a 32 % decrease in use [7] as well as a 23 % cost saving [87], with no significant adverse consequence to either length of stay or mortality.

Our own group, the surgical critical care initiative [33] conducted an internal cost-savings analysis of a tool meant to improve the medical treatment of open wounds associated with trauma, across the injury cycle (days in the intensive care unit and general ward, rehabilitation, avoidance of hospital-acquired infections); those estimates point to savings of $50,000 per patient, or potentially $3.5 billion annually if applied across the 67,000 patients being treated each year with severe extremity injuries (data not published).

Taken together, and enhanced with the observation that the ability of CDSS to prevent incorrect decisions has been repeatedly demonstrated in the literature, these figures support a strong argument that a single superior decision can have substantial and positive impact on healthcare economics.

The impact of integrating and processing large data sets to better inform decisions also has implications in terms of productivity. A 2015 Mayo Clinic study on the effect of decision support on the efficiency of primary care providers in the outpatient practice suggests that use of a CDSS can result in a 65 % estimated saving of the provider’s time [88]; in this 2015 study, Wagholikar et al. assigned thirty patients to ten physicians, half of whom were assigned a CDSS to decide on preventive services and chronic disease management, ultimately pointing to an estimated saving of 3 min per patient. While the study does not touch on this question, the increased efficiency likely had positive ramifications in other areas of clinical care (e.g. number of patients being seen, time at the bedside, research).

There are nevertheless a couple of cautions. In the context of ever-rising concerns about cost-containment, more research in this field is needed; given that only a handful of studies measure cost at all, and only one reports on cost-effectiveness [89]. Future analyses should report total costs transparently, including costs on the aggregate, from development, to maintenance.

No CDSS can be perfect, and this immutable fact raises the spectre of liability exposure in cases of harmful error; while raised [90], this latter issue has not been fully addressed. Good faith reporting of errors is essential. Vetting for clinical use can include methods such as decision curve analysis [91] and net reclassification improvement [92]. These methods can be used not only to compare multiple models, but more importantly to determine which model(s), if any, are ready for insertion and use in the clinical environment.

Regulatory concerns

While newer methods, like decision curve analysis, assist in determining clinical utility, the risks stemming from the use of CDSS still remain largely unknown. Historically, developers have been predominantly concerned with the efficacy of systems rather than their safety [93]. Important to the debate is the fact that the former is generally much easier to demonstrate [94]. As a result, the question of quality and legal liability has become increasingly relevant to the discussion.

The United States Food and Drug Administration [95], the body regulating medical devices, has not been idle in monitoring health information technology (HIT) malfunctions with the potential for human harm. In 2010, these types of malfunctions seemed to be involved in 44 injuries and 6 deaths. Because registration and reporting by vendors is voluntary, these numbers could greatly underestimate a much larger concern [96]. From a safety perspective such occurrences must obviously be interpreted in the relative context of lives saved and outcomes improved by the use of CDSS.

As reported by the FDA, adverse events tend to fall into four major categories. The first category consists of errors of commission, defined as accessing the wrong patient record or overwriting information. The second category consists of errors of omission or transmission, defined as loss or corruption of vital patient data. The third category consists of errors in data analysis. The fourth category includes incompatibility between multi-vendor application and systems, which not only ties into the aforementioned concerns but again speaks to the challenge of interoperability.

The use of notebooks, tablets, or phones as deployment platforms for CDSS introduces yet another layer of complexity. While the FDA provided recent guidance, releasing its guidance on mobile medical applications in 2015 [95], the proposed recommendations remain non-binding and the determination of whether a Medical Device Data System (MDSS) falls under either a high-risk (class III) or a low-risk (class I) taxonomy appears to be up to the users [97].

The FDA considers that both hardware and software tools may be regulated if they fit the definition of a ‘device’, broadly defined as a mobile application which intended for use in the diagnosis of disease or other conditions, the cure, mitigation, treatment, or prevention of a disease, or is intended to affect the structure or function of the body. The Guidance appears to exclude general-purpose calculators or platforms providing medical reference materials [98].

Developers of CDSS must recognize that regulatory oversight, as it pertains to medical devices, is guided primarily by function (i.e. the platform itself, while relevant, does not determine classification). As an example, applications performing simple calculations routinely used in clinical practice (e.g. Body Mass Index, National Institutes of Health Stroke Scale) fall under the ‘enforcement discretion’ category whereas those performing patient-specific analysis along with diagnosis and treatment recommendations must be regulated [98]. In the end, despite several guidance documents released in 2014 and 2015, the ambiguity of classifying CDSS remains, and with it the ability by both developers and providers to determine whether the tool requires a regulatory pathway.

It is apparent that the FDA is taking a growing interest in CDSS and, in the context of maximizing patient safety and implementing post-market surveillance, is cautiously moving towards establishing standards, as well as a process to aggregate and analyze trends to mitigate future risk. To the extent that such tools are required for precision diagnosis, there is some urgency in accelerating development of those standards.

Our conjecture is that CDSS will be increasingly developed and deployed on mobile platforms; the FDA releasing specific guidance on the issue lends credence to our observation. However, much work remains to define and categorize these mobile CDSS, and to subsequently develop recommendations, for review and endorsement by the regulatory bodies. Our community of developers and users should embrace the opportunity to shape this debate, developing and proposing its own sets of risk-mitigation guidelines.

Conclusion

Clinical decision support tools are clinically meaningful to providers and have a measurable impact on diagnosis, treatment and ultimately health of the individual patient. The challenges facing the clinical insertion of medical decision-making tools are however abundant, including integration of large and heterogeneous data sets, provider resistance, and regulatory compliance. Despite these hurdles, CDSS have the potential to effectively manage utilization and improve both efficiency and precision, particularly in disciplines requiring complex decision-making. However, these benefits will only be recognized through randomized controlled trials (i.e. demonstration of effectiveness) and standards sanctioned by regulatory bodies.

Of equal importance is the realization that, in the context of maximizing clinical utility, ‘next-gen’ CDSS need to go beyond population-based risk stratification and move towards individualized and precise risk-predictions. In so doing, these will contribute to the promise of a more personalized approach to medicine by delivering the right treatment, to the right patient, at the right time. Others in the oncology and cardiac transplant arenas have begun to make that leap; we invite the whole medical community to follow suit.