At least 44,000 people die in hospitals each year as a result of medical errors, and these deaths are becoming the eighth-leading cause of death in the United States. Thus, medical providers have the responsibility to pay attention for reducing avoidable medical errors and improve patient safety as best as they can. It requires the rapid evaluation and prioritisation of life threatening injuries in the primary survey followed by a detailed secondary survey in the emergency room. However, time is always valuable and limited such that some important vital signs may be delayed and ignored. This research explores delayed diagnosis problem and uses the affinity set by Topology concept to classify/focus on key attributes causing delayed diagnosis (missed injury) in order to reduce error risk. Results interestingly indicate that when a patient can breathe normally, but his (or her) blood-pressure or pulse is abnormal, a high probability of delayed diagnosis exists. This affinity work also compares the performance with the model of rough set (Rosetta), neural network, support vector machine and logistic regression. And our affinity model shows its advantage by prediction accuracy and explanation power.
Curing disease, maintaining health, and saving lives is the doctor’s mission. People in traditional society believe the doctor’s expertise is indisputable, trustworthy, and unmistakable. However, hospitals are not as safe as they should be, according to the results of two studies, the “Harvard Medical Practice Study” (Leape et al., 1991) and Institute of Medicine (IOM) investigation report “To Err Is Human” (Kohn, Corrigan, & Donaldson, 1999) published in 1986 and November 1999. The Harvard Study shows that 3.7% hospitalized patients suffered from medical injury, 27.6% suffered from medical negligence, 69% due to human error, 2.6% patients suffered permanent disability and 13.6% of medical errors led to death. On the other hand, the IOM report also shows that at least 44,000 people and perhaps as many as 99,000 people die in hospitals each year resulting from avoidable medical errors (Leape et al., 1991). These reports show that people dying in a given year due to avoidable medical errors exceed those of motor vehicle accidents (43,458), breast cancer (42,297), and AIDS (16,516) (Leape et al., 1991). Medical error deaths are becoming the eighth-leading cause of death in the United States (Kohn et al., 1999).
The aforementioned observations point out that reducing medical errors is critical when most are preventable (Department of Health). Diagnostic delay not only has clinical implications in terms of worsened outcome and potential long-term disability, but may also have financial and medico-legal consequences (Aaland and Smith, 1996, Brooksa et al., 2004 and Thomson and Greaves, 2008). Actually, real data of missed injury is hard to be defined and collected. Many researches indicate there is a strong link between the diagnosis delay and missed injury: when a patient is delayed, there is greater probability to generate medical errors (Born et al., 1989 and Furnival et al., 1996). Therefore, this study explores diagnosis delay or failure of a planned action in operation in order to reduce the error risk. Diagnosis delay means patients’ injuries are ignored or missed in emergency room (ER), but are identified by doctors in intensive care unit (ICU). The purpose of this paper is to find key attributes, which may lead to the delayed diagnosis problem by affinity set data-mining. The affinity set (Chen and Larbani, 2006 and Larbani and Chen, 2008) is inspired from the vague interaction between people in social sciences (Freeman, 2004, Ho, 1998, Hwang, 1987 and Luo, 2000), developed by Prof. Larbani and Prof. Chen as the data-mining tool to classify, analyze, and build the relationship between observed outcomes (consequences) and possible incomes (causes) of an information system. This research collects clinical data from ER in Chung-Ho Memorial Hospital, Kaohsiung Medical University, Taiwan; then uses the affinity data-mining model to identify key attributes leading to delayed diagnosis. Affinity mining results are also compared with some popular approaches for their performances, and our affinity model has the advantage over the model of rough set (Rosetta) (Pawlak, 1982 and Pawlak, 1991), neural network (Zbikowski & Hunt, 1996), support vector machine (Cortes & Vapnik, 1995) and logistics regression (Collett, 2003).
This paper is organized as follows: Section 2 introduces the basic concepts and definitions of affinity set; after that, the affinity data-mining model is proposed. Section 3 uses actual samples from Kaohsiung Medical University hospital, Taiwan to validate our affinity data-mining idea, deriving key delayed diagnosis attributes. This affinity work also competes with the model of rough set, neural network, support vector machine and logistic regression. Finally, Section 4 gives conclusions and recommendations based on our current achievements of the affinity data-mining model
This is a successful beginning of data-mining attempt by affinity set. Although the initial achievements are encouraging to doctors, many valuable problems still wait for resolution from this small beginning. For example, delayed diagnosis may be not directly/strongly related to the missed injuries or medical errors, because a patient vital sign is dynamically changed and not easy to be captured in time. Thus, clearly clarifying this point would be valuable to capture the real causes of missed injures. In addition, adding other attributes, such as workload of doctor, number of doctors in ER, available medical resources, patients’ arrival time, body trauma property, injury location and severity, etc., to analyze relations within medical data will enhance this research’s quality. Other mapping/projection methods inspired from Topology (Kelley, 1975 and Mendelson, 1990) may also generate effective rules for this research. In other words, the rule base V defined in this study is not unique and the only one. When the evaluation/generation of rules becomes a tough task, evolutionary algorithms: e.g., genetic algorithms may be valuable.
Readers should also recognize that: (a) the affinity model is quite simple, and (b) no fuzzy membership functions are assumed in this study; thus, we do not think our affinity data-mining model is fuzzy. Of course, integrating affinity set and fuzzy set may be valuable. Finally, this study validates the superiority of affinity model over other compared models in this study. And an efficient communication between patients and doctors is necessary/emergent in ER, e.g., building radio frequency identification (RFID) systems to capture all patient vital signs with clicks in seconds would be very beneficial.
