Personalized statistical medicine

The term ‘statistical medicine’ is generally used for most of the present-day medical practices of using group-based research results on individual cases1. For example, when a sufficiently powered clinical trial on a group of cases establishes that a new regimen B is better than the existing regimen A for treating disease X, and the difference is medically significant, the new regimen is adopted on nearly all cases of disease X. Such studies do have inclusion-exclusion criteria that specify the focus group, but the results are not personalized. On the opposite side is personalized medicine where the management of a case is based on his or her individual characteristics, particularly the genetic profile2. We present here a new approach to medicine where diagnosis and prognosis assessments of a patient are based primarily on statistical assessment of the individual characteristics and call this personalized statistical medicine. This communication, an extension of our previous proposal to consider statistical medicine as a new medical speciality3, provides a new personalized perspective, explains the elements of personalized statistical medicine and discusses how this can be practiced for the benefit of the individual patients without incurring much cost.

The present paradigm of statistical medicine

The term statistical medicine is widely used for application of a statistical approach to medical problems such as to understand the correlation between different medical parameters and health states4, to study the cause-effect relationships5, for assessing the quality of data6 or simply for biostatistical methods7. These collectively come under the gamut of, what is called Medical Statistics in the UK and Biostatistics in the USA. To signify its focus on medicine rather than veterinary and agricultural sciences, the term ‘medical biostatistics’ was suggested with an argument that ‘medical + bio’ content of the subject is more than the ‘statistics’ content8.

Most of the medicine practiced today is based on statistical results that show a regimen to be effective in a majority of cases of a specific type and is generally adopted for almost all cases of that particular type unless there is some unusual presentation. This approach may, however, not work in some cases. Guidelines for diagnosis, treatment and prognosis of diseases are usually established on the basis of the statistical results of clinical trials and such other studies9. This has motivated a flurry of research for new biostatistical methods, and new approaches of inferences are being developed to come to a reliable and valid conclusion10. However, each person is unique and the management of health of each is tweaked as considered appropriate by the concerned clinician to try to personalize the regimen to the individual need even in this approach.

Personalized medicine

A press release of the International Medical University in Rome considered it unusual that a mathematical discipline has a dominant role in the medical field11, and Fulweiler1 expressed complete disappointment with the current practice of clinical decisions to be based on statistical evidence. There is a pleading to do away with such statistical medicine and instead practice, what is generally termed as, personalized medicine9. This is considered opposite to the prevalent statistical medicine. There are numerous challenges in its way12 although many papers have come out recently on using personalized medicine for different diseases. The chance factor can still be prominent, and it is yet to be assessed how much of the chance element can be reduced by using targeted approach under such personalized medicine framework.

Personalized statistical medicine

An ailment is an aberration that derails the process of homeostasis and can arise due factors such as injury, pathogens, stress and degeneration. Treatment is an effort to put the body’s system back on track. Decisions regarding diagnosis, treatment, and prognosis are currently based on clinical assessment of the presenting signs and symptoms; history; patient’s characteristics such as age, sex and nutritional status; the environment such as the availability of medical facilities and the cost affordability. Help of clinical measurements and laboratory and radiological assessments is taken to provide support. Yet, medical science can hardly ever be deterministic since all the events cannot be entirely explained by its causes13.

The goal of personalized medicine is to manage a case on the basis of the person’s individual characteristics rather than on the results of the studies on groups of cases of that particular type. Probabilities and uncertainties remain prominent in this setup also. Everyday research in different medical disciplines is an effort to shrink the space for chance, where chance can be understood as the set of causes that are unknown or too difficult to comprehend8. There exists an additional avenue by way of statistical tools that can help reduce subjectivity and minimize the role of chance in day-to-day clinical decisions, although the role of chance can never be eliminated because of unpredictable human variation. This avenue is provided by scoring systems and such other statistical tools as per the following details that can help make more objective decisions.

Scoring systems, rating scales and indexes: Our review of literature suggests that more than a thousand scoring systems, scales and indexes are used for clinical decisions. Prominent examples for score are APACHE (Acute Physiology and Chronic Health Evaluation) score, APGAR (Appearance, Pulse, Grimace, Activity and Respiration) score and pain score, and for scales are the depression assessment scale, Glasgow coma scale and end-stage liver disease scale. Indexes in common use are body mass index, glycaemic index and perfusion index. These tools provide a composite measure of two or more characteristics and are generally perceived to provide more comprehensive and objective pictures of a particular aspect of the health status of a person. Although these tools are not as precise as the laboratory measurements, these are commonly used to assist clinicians in evaluating the condition of the patient, thus helping in establishing the diagnosis, prescribing the treatment and assessing the prognosis in much the same way as laboratory results do. Also the ever-increasing development and use of such tools indicates that these are helpful in the management of different health conditions.

Scoring systems can incorporate both quantitative and qualitative characteristics, such as ejection fraction on the one hand and dizziness on the other. For example, APACHE-II score, among others, comprises pH value, age category and the surgical history – pH is a metric, age is categorical and surgical history is a nominal characteristic. A large number of rating scales are commonly used in psychological evaluations. An index is mostly calculated by combining two or more metric measurements. For example, left atrial (LA) expansion index is calculated as the difference between maximum and minimal LA volume as a percentage of the minimal volume14. We all use body mass index based on weight and height.

Besides that such composite measures provide a comprehensive assessment and tend to introduce objectivity by minimizing the chance of bias, these are versatile too in their ability to combine metric, ordinal and nominal measurements. The disadvantage is that many such composite measures now in use lack scientific validation under varying conditions. Nevertheless, the advantages far outweigh the disadvantages.

Models for prediction of outcomes: Establishing a correct diagnosis is in the domain of prediction, so is the most suitable treatment strategy for any particular patient. A prognosis has future overtones and is quite often difficult to predict with a degree of assurance. While clinical acumen is asset, help of statistical models is welcome by many clinicians. Prediction models15 provide a direct tool for personalized medicine. Cho et al16 compared the performance of statistical models with machine learning (ML) methods for predicting readmission and mortality in patients with myocardial infarction, and Grove17 found that such tool-based ‘mechanical prediction’ on an average is 10 per cent more accurate than clinical prediction. More studies are, however, needed to compare the results of statistical predictions with clinical predictions, but inexpensive statistical tools seem to have a great potential in improved healthcare. Among the recently proposed are models to predict a diagnosis before bile acid determination18 and for differential diagnosis of bacterial and viral meningitis19.

Decision trees: This approach has been recently forwarded for pressure ulcer risk assessment in immobilized patients20 and breast cancer diagnosis21. These trees combine probabilities of different outcomes and the clinician’s value judgement at successive decision nodes to progressively assess the likelihood of various outcomes. Thus, a decision with the highest chance of success can be taken in the best interest of the patient.

Artificial intelligence (AI) and machine learning (ML): Of late, artificial intelligence (AI) and machine learning (ML) have made considerable in-roads into the clinical decision process. Both these are data-intensive methods and have substantial statistical content. Liscia et al22 have proposed AI for the diagnosis of Helicobacter pylori in gastric biopsies. Cross and Harding23 have discussed AI in risk profiling in the treatment of chronic wounds. Similarly, ML has been suggested to enhance clinical judgement in the care of high-risk heart transplant recepients24. Ahmed et al25 have reviewed ML approaches in the identification of paediatric epilepsy. The development and usage of these methods require a statistical data analytics approach for clinical decisions which this communication intends highlight. A summary of tools and examples of their clinical use are provided in the Table.

Other statistical tools for personalized clinical decisions

In addition to the tools presented in the Table, several other statistical tools are also used every day in clinical practice such as probabilities in diagnosis, treatment and prognosis, and reference intervals. These are not personalized but are directly used on individual patients in clinics. Some others, as listed next, have indirect use. These are well known, but their role in personalized medicine is less appreciated.

Probability in diagnosis, treatment and prognosis: When a diagnosis is established, it is the ‘most likely’ diagnosis for the observed clinical features. When a treatment is prescribed, it is the one ‘most likely’ to provide the greatest relief to the patient. Similarly, probability is used for prognosis as well. However, probabilities are a measure of belief and reflect how much we know about the event. Morgan et al31 observed that practitioners generally overestimate the probability of diagnosis before testing as well as after testing. They also commented that many practitioners are not accustomed to using probability in diagnosis and other clinical decisions. Training of doctors in more judicious use of probability may help in personalized care.

Reference intervals for diagnosis and prognosis: Reference intervals of medical parameters have a dominant role in medical practice. Whereas some intervals are clinically determined such as blood pressure ≥140/90 mmHg for hypertension and fasting blood glucose level ≥126 mg/dl for diabetes, for most other parameters, such intervals are statistically determined as 2.5^th to 97.5^th percentiles of the values seen in healthy persons. These intervals are used not just for establishing diagnosis but also to monitor the progress of the patient and for prescribing and calibrating the treatment regimens.

Clinical evaluation of the validity of investigation results: Zhou et al32 discussed the role of statistical evaluation of medical tests that are extensively used for establishing the diagnosis. This evaluation is typically done in terms of different validity measures such as sensitivity, specificity, positive and negative predictive value and the area under the receiver operating characteristic (ROC) curve. ROC curve is also used to locate the cut-off of a quantitative test that best discriminates subjects with and without a condition. This cut-off is used in individual cases for optimal results. Sensitivity-specificity/predictivities have become an integral part of medicine and their role seems to be increasing by the day as the demand for objective assessment is increasing.

Relative risk (RR) and odds ratio (OR) for quantifying the risk: Relative risk (RR) and odds ratio (OR) are among the most common statistical indices used to assess the risk of an outcome when a purported risk factor is present vs. absent. Whereas their statistical significance is considered as evidence of their association with the outcome, the values of RR and OR, when calculated for several risk factors, also quantify their relative importance for use in personalized care although this feature does not get enough attention. Multivariable logistic regression helps in obtaining the ‘net effect’ when other factors in the model are kept constant and helps in removing their confounding effect to arrive at more valid results.

P values and confidence intervals (CIS) for personalized care: A clinical trial is considered to provide sufficient evidence for the superiority of a regimen when its performance exceeds a clinically unimportant threshold and is significant statistically, indicated by a sufficiently small P value. However, a judicious decision is needed after considering the specific condition of individual patients, including the statistical deviation of the biomarkers from the desired values and risk-benefit assessment for a particular patient.

The CIS give a plausible range that is likely to include the parameter value. If the CI for the efficacy of a regimen is wide such as 53 and 83 per cent, one will know that this result is imprecise and alerts one to rely less on it for its clinical application.

Overall, the efforts to manage a medical case yield varying results, and uncertainties remain prominent to its disposition. For proper results, it is necessary to reduce this window of uncertainties. Statistical tools play a vital role in minimizing and assessing the uncertainties and can help in personalizing clinical medicine. Prominent among such tools are properly validated scoring systems, scales, indexes, prediction models and decision trees. These already exist for many conditions, but more and better statistical tools can be developed for increased validity and the reliability of the clinical results for application to a wider variety of medical conditions in individual patients.