The results show that the PredictAD tool alone (72%) and the clinician with the assistance of the PredictAD tool produced comparable or higher accuracy in predicting 3-year MCI outcome than current research criteria for diagnosis of prodromal AD. The literature is somewhat confusing, due to differences in size of study populations, statistical methods, and length of follow-up etc., but it seems that the overall accuracy of combinations of clinical data and/or biomarkers in predicting AD conversion from MCI has varied from 67% to 93% 
. Liu et al., using the 100 MCI cases from AddNeuroMed data and a combination of neuropsychological tests and structural MRI biomarkers reported overall accuracy 69% during one year follow-up 
. Studies with the ADNI cohort reported accuracies 67–77% when using combinations of clinical measures and CSF and MRI biomarkers 
We acknowledge that the prediction accuracy of about 70% is not high concerning the clinical utility but the result is still comparable with the current state-of-the-art. It reflects a reality that the current prodromal AD guidelines and combinations of biomarkers are not perfect. However, our point was not to develop a novel method but to show how the current guidelines compare with computer-assisted methods. The PredictAD tool can provide objective and evidence-based information about the state of the patient by integrating heterogeneous measurement data acquired from a patient in current clinical practice. PredictAD makes it possible to assess the disease severity, i.e. it is not simply a yes/no diagnosis. Itś graphical user interface can make it easy for clinician to explore every single test or biomarker, giving more confident to clinicians than a probability or yes/no diagnosis calculated with certain software with underlying complex statistical calculation. Using the PredictAD tool, the clinician was able to detect a sub-population for which the accuracy was 84% which starts to be high enough for affecting the clinical reasoning. It is good to remember that 100% is not the correct target value in reality due to different reasons: 1) Stable MCI and progressive MCI cases in ADNI are not pathologically confirmed cases. It has been shown in different studies that the agreement of the clinical and neuropathology diagnoses is 70–90% 
. In other words, even 72% is within this range and studies reporting values >90% should be interpreted with a caution. 2) Even neuropathological diagnoses are not perfect.
It is interesting that about 30% MCI cases with clear (DSI ≥0.83, 5 of 16 cases) and probable (0.67≤ DSI <0.83, 20 of 71 cases) indications of AD did not convert to AD during the 3-year follow-up, even though they did not significantly differ from AD converters in age, gender, presence of APOE4, years of education, concentrations of CSF markers, Scheltens scores, MMSE, and RAVLT results (). The reason why those 25 stable MCI cases did not convert to AD is still unknown. In fact, this subgroup population seems to be interesting, and a detailed investigation of this subgroup, we might uncover novel preventative factors which delay the onset of symptoms of AD.
Screenshots from the PredictAD tool for two cases.
Current research criteria for prodromal AD 
emphasizes that the core criteria of episodic memory impairment should not only include deficit on delayed free recall but also on cued recall or recognition. In this paper we used RAVLT free recall and recognition scores to form the criteria of episodic memory impairment. Adjustments for gender or education were not used, and in addition it can be argued that results may have been different if another memory test or cut-off values would have been used. However, it is essential to remind that all MCI subjects in ADNI cohort already fulfilled a significant memory impairment measured with WMS-R logical memory II test (with education correction). Thus subjects who fulfilled the criteria of episodic memory impairment in the present paper performed lower than expected for age altogether in three memory tests.
It has been shown that the Scheltens scale can classify AD patients and healthy controls or other types of dementia with high sensitivity, specificity, and accuracy 
. Westman et al. 
applied Scheltens scale 2 and 3 as cutoff values in 101 MCI cases from the multicenter study AddNeuroMed study. They reported that the visually evaluated atrophy of MTL produced similar accuracy in predicting conversion from MCI to AD (68%) compared to multivariate regional MRI classification and manual hippocampal volumes at one year follow-up. We applied Scheltens scale 3 as the cutoff value in the ADNI data and found prediction accuracy (62%) during the 3-year follow-up.
In the present study, according to the most recent criteria for likelihood of AD, only 25 cases fulfilled the high likelihood of AD, i.e. all clinical core criteria, MRI and CSF markers were positive, fifteen of those 25 (60%) cases did convert to AD. Moreover, very low sensitivities (6%–57%) were achieved by using the combination of clinical core tests, and MRI and CSF markers. In contrast, by using the PredictAD tool, the number of clinician’s diagnosis of a clear indication of AD was 62 cases, and 45 of those 62 (73%) cases did convert to AD. This finding indicates that the PredictAD tool uses the clinical, MRI, and CSF data in a much more efficient way than the recent criteria applied with specific cut-off values for making the diagnosis of AD.
We acknowledge that the present study has certain limitation. In the predicting AD conversion with current prodromal AD guidelines, only RAVLT tests were used to define if the subjects had prodromal AD symptoms, but in the predicting AD conversion with PredictAD tool alone, all the neuropsychological tests were used. When the clinician determined if the subjects had prodromal AD symptoms with the assistance of PredictAD tool, only RAVLT tests were used as in the predicting AD conversion with prodromal AD guidelines. However, the clinician was not blinded to the other neuropsychological tests, the performance at the other tests exploring cognitive domains other than memory were used to increase the confidence of clinical diagnosis. The overall predicting accuracy was 72%, 71%, and 64% for the PredictAD tool alone, clinician’s prediction with the assistance of the PredictAD tool, and the best combination of the core clinical and biomarkers respectively. Diagnosis with the PredictAD tool was significantly more accurate than diagnosis by biomarkers alone or the combinations of clinical core criteria and biomarkers. The methods judging if a subject presented prodromal symptoms were not equal. It may explain the differences in overall predicting accuracy. The findings imply that a single neuropsychological test is not powerful enough to replace the other neuropsychological tests in early AD diagnosis, enhancing the justification of using PredictAD tool in clinical practice.
In conclusion, with the assistance of the PredictAD tool, the clinician can predict AD conversion more accurately than than the current research criteria for prodromal AD.